Changelog

All notable changes to this project are documented in this file.

The format is based on Keep a Changelog; version numbers use Semantic Versioning notation but are bumped patch-only during 0.x (each release +0.0.1; no minor/major jumps).

Every entry is numbered in one global per-category sequence — Decision, Change, Feature, Bug — and split across the version it shipped in. Within a version the list is sorted Decision › Change › Feature › Bug, each category newest-first (highest number on top). Numbers are never reused. (Open/unsolved items and the hard-won lessons live in docs/development/issues.md.)

[0.6.15] - 2026-07-22

The last hand-rolled embed serialization is gone — "Replace image" / "Replace all" now build through the same ONE grammar writer as every other rewrite path; the lint:obsidian 0-error baseline is restored; and four post-processor-hosted-embed bugs (invisible/duplicated table-cell renders, a filename shown as caption, no hover toolbar) are fixed as one host-agnostic attach decision instead of four separate patches. A second batch removes the plugin's LAST automatism: the passive link/size normalizer is gone (new F0/F27 — writes only on explicit user action), every ordered edit now rewrites its embed canonically instead, and the table-host defect class the normalizer's rewrites kept re-triggering (dead cells / raw {…} shown) is fixed at the attach/render-lifecycle level — covered by a permanent full-matrix guard (verify-table-host.mjs, 60/60; npm test 379). A third batch merges four small string/regex formulas that had drifted into duplicate copies onto their one shared source (crop editor ↔ render-core's frame-orientation string, three copies of the CSS-function-string regex, render-core's foreign-page flip/length reader ↔ transforms.ts, and the standalone runtime's caption gaining Bug 121's filename-as-caption guard) — npm test 385.

• Change 60 — The standalone runtime's caption now guards against alt being just the image's own filename (the off-Obsidian counterpart of Bug 121). What: a new pure captionFromAltGuarded(alt, srcOrFilename) in caption-logic.ts wraps captionFromAlt and returns "" when the parsed caption is EXACTLY the image source's basename (a local decodeURIComponent + last-path-segment implementation, not an image-resolver.ts import — that module carries a type-only obsidian import the runtime bundle must never pull in, T3); runtime.ts's addCaption now calls it with img.getAttribute("src") ?? img.src. The plugin itself already avoids this failure class architecturally (reading captions from the SOURCE text, never the rendered alt, since Bug 121) — the runtime has no source text to fall back on, so it needed the cheap heuristic guard instead. Why: R0 — closes the one remaining instance of the "filename shown as caption" defect class that the plugin already solved elsewhere, without coupling the Obsidian-free runtime/caption-logic modules to the plugin's source-resolution machinery. Unit-tested (exact-filename match incl. path prefix + query string, %20-decoded match, a real caption that merely contains the filename, a real unrelated caption, size-only/empty alt) in caption.test.ts; implementation-plan.md updated (gated).
• Change 59 — render-core's readTransform (foreign-page attribute reader) reuses transforms.ts's flip-token vocabulary and bare-number-to-px rule instead of its own copies. What: readTransform's inline flip-token loop (h/horizontal/v/vertical/both) is replaced by a call to a newly-exported parseFlipTokens(v) — the exact same tokens parseAltText's flip= case now also calls; its inline width/height bare-number-to-px regex is replaced by a call to transforms.ts's now-exported lengthValue (verified semantically identical: both trim, both test /^\d+(\.\d+)?$/, both append px only for a bare number — no behavior change). Why: R0/DRY — the foreign-page reader and the {…}-block parser read the identical bare-key grammar and had drifted into two copies of each rule. Unit-tested (parseFlipTokens, lengthValue direct cases) in transforms.test.ts; implementation-plan.md updated (gated).
• Change 58 — The CSS-function-string regex (/([a-zA-Z][\w-]*)\(([^)]*)\)/g) merged from three copies onto transforms.ts's one. What: transforms.ts's private parseFns is now exported; export.ts's parseTransformFns and crop-editor-logic.ts's parsePlacement both had their own byte-identical regex literal — both now call the shared parseFns and split its args string on , themselves at the point of use (their per-arg needs differ: export.ts wants an array up front, crop-editor-logic parses it lazily per matched function), instead of re-scanning the string with their own copy of the same regex. fnsToString stays private (no external caller needs it). Why: R0/DRY — the one CSS-function grammar had drifted into three independent literals. Unit-tested (parseFns on a multi-function string, empty/undefined input) in transforms.test.ts; the existing parsePlacement round-trip test (crop-editor-logic.test.ts) and renderTransformedImage export path stayed green unchanged, confirming behavior parity; implementation-plan.md updated (gated).
• Change 57 — The crop editor's frame-orientation formula merged onto render-core's (one string builder, not two). What: render-core.ts's applyOrientation now calls an exported pure orientationTransform(deg, flipH, flipV): string (the translate(-50%,-50%) [rotate] [scaleX(-1)] [scaleY(-1)] string); crop-editor.ts's private orientationTransform() (which duplicated the same four lines verbatim, per its own comment "the SAME string render-core writes") is gone — the editor now calls the exported function directly with its live (not-yet-committed) orientDeg + the existing transform's flipH/flipV. Why: R0/DRY — one formula, not a hand-kept copy the editor's own comment already flagged as a must-stay-identical duplicate. Unit-tested directly (identity/rotate/flip/combined cases) plus a source-text guard asserting crop-editor.ts imports the function and carries no const parts = ["translate(-50%, -50%)"] of its own, in render-core.test.ts; implementation-plan.md updated (gated).
• Change 56 — Passive normalization REMOVED (F0/F27); every ordered edit writes its embed canonically instead (F5/F6). What: the editor-change → 400ms normalize hook and both passes (normalizeLinkFormat, normalizeNativeSizes, main.ts) are deleted, and convertEmbedLine with them (link-format.ts) — the plugin never rewrites a line it was not ordered to edit, and a change of Obsidian's "Use [[Wikilinks]]" setting triggers no plugin action (the {…} syntax is form-independent). Instead the ordered-edit writers — writeTransform, the resize handle's rewriteWidth (live-preview-logic.ts), Replace — rebuild the WHOLE embed canonically: the form desiredFormat dictates, the path token from fileManager.generateMarkdownLink parsed by the one grammar (pathFromGeneratedLink — the generator returns a plain link without the embed !), a #/^ subpath stripped for resolving and re-attached as written, the native |size folded into the block, caption kept; canonicalTarget keeps the SOURCE form and path when no verified token exists or a wiki alias would have to carry ]] (never lose the link). An ordered edit that changes nothing writes nothing (a semantic no-op guard — an unchanged panel close adds no undo step and no canonicalisation-only rewrite). Why: the user's decision — no unasked writes, no background restructuring; the normalizer's background rewrites were also the trigger feeding the Bug 126 render window. Requirements F0/F5/F6/F27 rewritten by the user; unit tests moved onto the writer pipeline (link-format.test.ts, live-preview.test.ts); journey-verified live (CDP): a size commit writes the canonical line in ONE write and an unrelated edit afterwards triggers no follow-up write (editor-change counter 1→2, only the unrelated one).
• Change 55 — The Replace writer merged onto the one grammar writer (buildEmbed) — table-pipe escaping and the write ⊆ read invariant now cover Replace. What: buildReplacementEmbed/replaceEmbedTarget/planReplaceAll (replace-logic.ts) no longer assemble ![[…]]/![…](…) strings by hand — they call link-format's buildEmbed (Bug 120), so a table row's pipes are escaped (ImageLocation.inTable), an md destination is percent-encoded, an md caption's \/[/] are escaped, and a native size already on the embed folds into the {…} block like any other active edit (Bug 94 precedent, F6/T2) instead of being re-emitted as a raw pipe suffix. An embed whose caption cannot be represented in the target form (a wiki alias containing ]], which a wikilink has no escape for) keeps its EXISTING form on Replace — only the path swaps, never lose the link — mirroring convertEmbedLine's own policy. Why: R0/DRY — the last duplicate embed-string writer is gone; Replace now gets the same escaping/invariant guarantees for free instead of maintaining its own, incomplete copy. Unit-tested (table-row escaping, the ]]-caption keep-form fallback, size-fold parity with existing cases) in replace-logic.test.ts; implementation-plan.md updated (gated).
• Bug 130 — A reading-view caption RODE Obsidian's re-used embed DOM into live preview and showed as a stray visible caption on the suppressed native. Obsidian caches and re-parents embed views across reading/LP; a caption the reading adapter attached travelled with the DOM and nothing removed it (the native it rides is not this pass's render target). Fix: the AD5 suppression now covers the WHOLE native render — .cm-content .internal-embed.image-embed > .lie-caption is hidden, with a cascade re-allow inside .markdown-rendered render blocks (table cells keep theirs) (styles.css); the reconcile pass additionally removes a stale caption from a skipped CM6-native host (cache hygiene, F2), and applyReadingCaption's reuse check requires the previous caption to still be connected. CDP-verified in the verify-table-host caption passes.
• Bug 129 — After a mode switch the reconcile pass processed the HIDDEN view: reading view lost its captions and replacement boxes. reconcileFromSource picked its container with a bare selector list (.markdown-reading-view, .markdown-preview-view, .markdown-source-view) — querySelector returns whichever sits first in the DOM, which after switching modes is the hidden one, so the visible view was never walked. Fix: the container is picked by view.getMode() (main.ts). CDP-verified: reading-view captions render again after LP → reading.
• Bug 128 — "Replace image/all" truncated a parenthesis-bearing generated path at the first ). replacementPathToken extracted the md token with \]\(([^)]+)\) — ![](Screenshot%20(1).png) came back as Screenshot%20(1. Fix: the token now rides the one grammar (pathFromGeneratedLink → the scanner's balanced-paren destination parse); fallback stays the vault-relative path. Unit-tested in link-format.test.ts.
• Bug 127 — The canonical path token NEVER parsed: generateMarkdownLink returns a plain link (no embed !), so formattedPath always fell back to the raw written path; #/^ subpaths never resolved and were silently at risk. parseEmbedLine requires the !, so the parse of the generator's output failed on every call (verified live: gen("sample-portrait.png") → "[](sample-portrait.png)") — every link-form conversion shipped the UNVERIFIED raw wiki inner as an md path and only worked by coincidence of the vault's "shortest" setting; a ![[img.png#page]] target never resolved (no subpath strip before getFirstLinkpathDest), and exportImage had the same missing strip plus a dead path.split("|"). Fix: pathFromGeneratedLink (link-format.ts) normalizes the generator's shape before parsing with the one grammar; canonicalPathToken (main.ts) strips the subpath for resolving and re-attaches it as written (T12), and exportImage strips it too. On a null token the edit keeps the source form and path (never an unverified path — write ⊆ read). Unit-tested (bang normalization, parens, titles, format mismatch).
• Bug 126 — Table-cell embeds went DEAD after an edit in the cell: image gone or the raw {…} shown, deterministically toggling with the cursor (the user's lead report). Cause (CDP-traced): opening a row's cell editor makes Obsidian REBUILD the static cell copy hidden; the attach path decided finally on that render — a detached section was concluded early (one-frame recheck), a hidden copy was skipped as "retired" (Bug 122's model) — and NOTHING re-evaluated when the cell editor closed and re-showed the copy (closing is not a render), so the AD5 suppression stood without its replacement: an empty cell, or the un-stripped {…} beside the image. Fix (attach rides real render lifecycle only, F27 — no observers, no polling, no hover-triggered work): the post-processor attaches at the section's MOUNT (ctx.addChild → MarkdownRenderChild.onload); a copy rendered hidden is decorated at ITS render (its decoration shows nothing until Obsidian unhides it); the CM6 ViewPlugin create/destroy — the LP extension runs in every editor, the cell editor included — schedules the reconcile (the close covers the re-shown copy); the reconcile strips a leftover {…} text node via the shared strip helper and owns hosts structurally (anything in .cm-content outside a .markdown-rendered render block is CM6-native and skipped — the old adjacent-sibling .lie-wrapper guess broke on CM's widget buffers). Guarded permanently by verify-table-host.mjs: the full host matrix ({wiki, md} × {plain, size, block, size+block} × {LP static, real cell click, cursor out, reading} × both link-format settings × reveal/caption passes) plus the real size-modal journey, with instrumentation proving hover triggers ZERO reconcile runs and the counter settles when idle — red on the old build (dead cells reproduced), 60/60 green twice on the fix.
• Bug 125 — lint:obsidian baseline broken: 2 errors where CLAUDE.md's gate requires 0. Two errors with DIFFERENT origins had crept past the last "0 errors" check (2026-06-27) unnoticed: src/live-preview.ts:514 @typescript-eslint/no-floating-promises — the AD10 reparse-rebuild's Promise.resolve().then(...) dispatch, introduced by the raw-link reveal rework and shipped in v0.6.12; and src/runtime.ts:95 obsidianmd/no-global-this — the Bug 119 shim itself (Object.assign(globalThis, { activeDocument: document, activeWindow: window })), added by the v0.6.13 release. Fix (two minimal, behavior-preserving code changes; no config touched): the reparse-rebuild statement is prefixed with void (the .then() callback already catches internally, the operator only marks the unconsumed promise as intentionally ignored), and the shim assigns onto window instead of globalThis — identical in this browser main-thread bundle (window === globalThis, bare activeDocument/activeWindow still resolve) and exactly the alternative the rule's own message names; the guard regex in runtime-global-shim.test.ts now accepts both semantically identical forms (the guarded invariant — shim present and ahead of hydrate — is unchanged). npm run lint:obsidian back to 0 errors (18 documented warnings), npm test 347 green.
• Bug 124 — Plain embed in a table cell rendered INVISIBLE: the uniform native-suppression CSS hid the img with no replacement attached. processBlock's attach only built the plugin's 3-layer box when the {…} block carried a real transform — a parameter-less embed nested inside live preview's .cm-content (a table cell) stayed a naked native <img>, which the unconditional suppression rule then hid with nothing to show. Fix: attach now also builds the (degenerate) box whenever the host sits where that suppression applies (main.ts processBlock), re-validated one frame later since the section may still be detached at the synchronous post-processor call — suppression and its replacement are one decision, never one without the other. CDP-verified live: a parameter-less ![[img]] table-cell embed is visible again.
• Bug 123 — No HOVER toolbar on table/callout/footnote-hosted embeds — only a click opened it. The delegated mouseover only recognised .lie-wrapper.lie-float hosts; a post-processor-rendered embed has no such wrapper. Fix: the same delegate now also opens the floating toolbar for any .internal-embed.image-embed.lie-embed host (the marker attach sets on a host it decorated), reusing the identical region-hover binding the float path already uses (main.ts). CDP-verified live with a real pointer hover (Input domain, no click): the floating toolbar opens on hover and dismisses after the travel grace once the pointer leaves.
• Bug 122 — Cursor into a table row DUPLICATED the embed render (stale hidden copy plus the live cell editor's own, both decorated). Obsidian hides the table cell's static render (display:none on .table-cell-wrapper) once its row's own live cell editor takes over, but attach never recognised the hidden copy as retired. Fix: a small pure predicate, isHiddenHostCopy (attach-logic.ts, unit-tested in attach-logic.test.ts), makes attach skip a host copy Obsidian has already superseded and hidden — idempotent per host copy. CDP-verified live: exactly one visible image per cell with the cursor in the row (down from the stale-plus-live double decoration).
• Bug 121 — The rendered filename shown as CAPTION for every plain embed in reading view and in post-processor-hosted embeds (table cells, …). applyReadingCaption took the DOM alt attribute at face value — Obsidian defaults that to the bare filename for an embed without an alias, which is not author text. Fix: the caption now derives from the SOURCE text via the existing position-exact resolver (findImageInText/occurrence), exactly like the live-preview widget already does — reconcileFromSource's document-wide occurrence walk was extended to also reach post-processor-hosted embeds nested in live preview (refined its skip condition off the .lie-wrapper sibling the CM6-native embed always carries, rather than a blanket .cm-editor skip), and processBlock's own per-section attempt was dropped once ctx.getSectionInfo was confirmed (CDP) to return null for a table-cell host — one caption source for both adapters. CDP-verified live: a plain embed shows no caption, an aliased one shows its real caption, both in reading view and in table cells.

[0.6.14] - 2026-07-22

The embed grammar completed by design — a hand-written scanner replaces every embed regex in the project, so reading an embed matches everything Obsidian's own parser accepts within Markdown syntax and writing emits only Obsidian's canonical form; the auto-normalizer's own size-fold path rides the same grammar round-trip. Verified live (CDP); npm test 342.

• Change 54 — The md native-size fold rides the ONE grammar round-trip; a conflicting width=/height= block key is replaced, not duplicated. What: the auto-normalizer's fold (normalizeNativeSizes, main.ts) drops its own embed regex (incl. its private \| table-escape knowledge) and parseAltText/serializeTransform rebuild — per line it now rides the same parseEmbedLine → buildEmbed round-trip as every other rewrite (only an md embed with a native size folds; a wikilink's native size stays in place, F5/F6). foldSizeIntoBlock (link-format.ts) now REPLACES a width=/height= key already in the block for each axis the native size sets (the native pipe size wins) instead of appending a duplicate key — quote-aware, every other token keeps its place. Why: R0/DRY — the same fold existed twice, with duplicated escape knowledge outside the one grammar source. Unit-tested (conflict replacement, height axis, foreign-token order) in link-format.test.ts; implementation-plan.md updated (gated).
• Bug 120 — The embed grammar ignored Markdown's escape/encoding layer — table embeds were destroyed on open (normalizer data loss), size/crop submenus muted, \|<size> unfoldable, and paths like Screenshot (1).png unparseable. The trigger was the table-escaped pipe: in a table row a content pipe is written \|, but the grammar split the wiki inner at the first RAW pipe, so ![[p\|90]] parsed with path p\ — merely opening a note let the auto-normalizer rewrite the row to a broken ![](p\){width=90} (no image, raw attr text, deterministic data loss), locateActiveImage never matched such embeds (size/crop submenus silently dead on those rows), and a resize couldn't remove the native \|90. The insight is general — Markdown links carry a whole escape/encoding layer, and any editor may write it — so the fix is the LAYER, by design, at the one grammar source (link-format): a hand-written scanner replaces every embed regex in the project (the resolver's and live-preview-logic's included). READ accepts what Obsidian's parser reads within Markdown syntax (verified live against metadataCache): a wiki inner to the first ]] (single [/] legal), the table layer stripped (\| ≙ |), the alias at the first pipe, a #/^ subpath split off for resolution; an md alt with CommonMark backslash-escapes; an md destination bare with arbitrary-depth balanced or \(\)-escaped parens (unbalanced → not an embed), as <…> angle form, or with a trailing "…" title (recognized, discarded); %-decoding for comparisons. WRITE emits only Obsidian's canonical form: md destinations percent-encode exactly Obsidian's set (space/backslash/control), the angle form is never newly produced, wiki stays raw, and into a table row every pipe goes out escaped (escapePipe, carried via ImageLocation.inTable). Verified live (CDP): opening no longer rewrites \| rows destructively (wiki mode leaves them byte-identical; md mode folds them into valid {width=…} embeds), the size submenu opens on \| rows, and a width commit folds \|90 → {width=75} cleanly.

[0.6.13] - 2026-07-15

Hotfix for the standalone runtime: lie-runtime.js crashed on every foreign page since 0.6.9 because the Change 40 sweep leaked the Obsidian-only activeDocument global into the shared core the runtime bundles. Verified: npm test 306 (incl. the new 3-case shim guard), npm run build green, built lie-runtime.js contains the shim.

• Bug 119 — Standalone runtime threw ReferenceError: activeDocument is not defined on every claimed image (broken since 0.6.9). Cause: Change 40's eslint-driven document → activeDocument sweep (0.6.9) deliberately skipped runtime.ts but converted the exact flagged positions in the SHARED core too — ensureLayers in render-core.ts and createCaptionEl/mountCaption in caption-dom.ts — which the runtime bundle also pulls in; off-Obsidian the activeDocument global does not exist, so the first hydrate crashed and no image rendered. No test covered the DOM path (the render-core unit test only asserts CLAIM_SELECTOR). Fix: the runtime ENTRY (runtime.ts) now SHIMS the Obsidian-only globals — Object.assign(globalThis, { activeDocument: document, activeWindow: window }) at the top of run(), before the first hydrate. This is AD9's runtime-exception clause: off-Obsidian the runtime supplies the missing platform binding itself, exactly as it supplies its own inline-Markdown renderer. The shared core keeps referencing activeDocument unchanged, so both callers stay identical AND Feature 39's still-open window → activeWindow half is safe in the runtime too. A source-text guard runtime-global-shim.test.ts asserts the shim stays in the entry and ahead of hydrate. Pitfall recorded in implementation-plan.md (gated).

[0.6.12] - 2026-07-02

Closes out the reveal/dismiss rework by removing the last behaviour forks: the <> dismiss becomes per-embed and link-only (two embeds on one line dismiss independently, inline embeds included), the editing toolbar is unified into one definition shown in two presentations (in-chrome + floating) so the <> and every control behave identically everywhere, and clicking any image — body or resize handle — moves the caret onto that embed uniformly (pure hover never does). Closes Bug 104. Verified: npm test 303, verify-reveal-matrix 25/25, verify-resize-affordance 9/9, verify-reveal 5/5, verify-optical-chrome 9/9, verify-crop 19/19.

• Change 53 — The editing toolbar is ONE definition shown in two presentations; the <> reveal is a shared item (closes Bug 104). What: buildToolbarElement renders the bar from the shared ToolbarItem[] for BOTH the in-chrome presentation (hosted by the EmbedWidget on the image box) and the floating one (ImageToolbar on body) — identical buttons, order and behaviour, only the host and a lie-toolbar-in-image / -floating class differ (toolbar.ts, live-preview.ts). The <> reveal, previously built OUT of the item model by makeRevealButton and prepended ONLY in-chrome, is now a normal leftmost ToolbarItem built in toolbarItemsForImage (main.ts): its action is the exported toggleEmbedReveal(img) (which recovers the editor via EditorView.findFromDOM and toggles the per-embed key), and its per-show is-off + label flip read the wrapper's .lie-dismissed class (a plain DOM signal set in the widget). makeRevealButton + the manual prepend/separator are gone; ToolbarButton gains an optional className. Why: the old fork meant a floated / inline (.lie-float) image — whose in-chrome bar is hidden — had NO reachable <> at all (Bug 104). One toolbar model makes the control present and identical in both views. Aligns the code with the already-recorded AB10 ("Toolbar", one) and D1.1 ("show the SAME bar floating"); implementation-plan.md updated (gated).
• Change 52 — The <> dismiss is now per-EMBED and link-only, superseding the per-line suppress (Bug 65 / Change 48). What: the dismiss is keyed by the embed's document position (e.attrEnd), NOT the line, so two embeds on one line dismiss independently and an inline embed dismisses too (the old e.mode !== "inline" exclusion is removed). Suppression is link-only: a per-embed lie-suppress-native MARK over the body span (e.from…e.embedEnd) hides only THAT embed's native raw link — no whole-line .lie-dismissed class, no !important — while the stand-in + {…} hide via their own withheld lie-show, so a sibling embed or surrounding text on the line is untouched. reduceReveal gains a lineOf mapper so the per-mode auto-clear maps each dismissed embed key back to its line (live-preview-logic.ts, live-preview.ts, styles.css). Why: Bug 65 / Change 48 suppressed the WHOLE cm-line, which couldn't target one of two embeds on a line and skipped inline embeds; keying per embed span realizes AD11 ("the plugin owns source-link visibility per embed span") faithfully. The .lie-dismissed class now rides the WRAPPER as a plain per-show DOM signal for the toolbar (no suppression).
• Bug 118 — Clicking an inline (mid-text) image didn't move the caret onto its source. Cause: two standalone-centric assumptions. (1) The click target was the LINE START (lineAt(posAtDOM(wrapper)).from) — for a standalone the embed IS the line start, but for a mid-text inline image the line start is other text, not the link. (2) The resize handle (a fixed-size marker half over the image corner) fully COVERS a tiny inline image, and the mousedown handler skips .image-resize-corner, so the click was swallowed and the caret didn't move at all. Fix: a shared moveCaretToEmbed(view, wrapper) lands the caret on the EMBED span (posAtDOM(wrapper)) uniformly for every mode, driven on the CLICK (pointerdown / mousedown) of BOTH the image body and the resize handle — so every active interaction moves the caret while pure hover moves nothing, and a plain click on the handle no longer resizes (a >2px drag threshold distinguishes a resize from a click) (live-preview.ts). CDP-verified: standalone + inline clicks land on the embed and reveal; the resize drag still rewrites the width.

[0.6.11] - 2026-06-27

The raw-link reveal rework — an image's source reveal/dismiss now derives from Obsidian's OWN parse (AB16b / D16 / D17) instead of a parallel regex or DOM guesswork, with a single native / auto / always reveal-mode dropdown (Feature 33) replacing the old boolean; the bare/inline embeds are unified onto the one uniform render, a reveal flip no longer rebuilds the image+caption (no flicker), the floating toolbar is reachable on cropped/short images, and inline is clarified as the vertical-alignment layout axis. Closes Bug 65 / 86 / 114. Verified: npm test 301, verify-reveal-matrix 25/25, verify-resize-affordance 9/9, plus the existing suite green.

• Decision 34 — Inline redefined as the VERTICAL-alignment axis, orthogonal to the five horizontal layout states and to size (doc-only). What: Clarifies the layout model in requirements.md: F15 and the retitled F17 ("Inline = the VERTICAL-alignment layout state") recast inline as the sole VERTICAL axis — it governs only an in-line image's vertical position on the line — orthogonal to the five HORIZONTAL block-left/centre/right · float-left/right states (which place an image taking its own block space) and orthogonal to SIZE (F24). The six F15 states therefore present as one mutually-exclusive choice while operating on two different axes (five horizontal, one vertical). F24 is clarified so icon sets the icon width (the smallest preset width), dropping the stale "sets the inline size (F17)" phrasing; the implementation-plan AB14 row for size-submenu-logic.ts is synced to describe icon = a line-height height and to drop the retired inline field from SizePreset, stating SIZE is orthogonal to layout. Why: Documentation catching up to the already-shipped "decouple size from layout" refactor (commit ba2dfa4) so the F-items and the plan match the code. Doc-only — no behaviour change: no layout-state mapping or size-preset code changed in this batch (the icon = line-height-height / SizePreset decoupling already shipped).
• Decision 33 — One engagement predicate isEngaged centralizes every "is this image active?" check (AD12). Added the single pure predicate isEngaged (toolbar-region-logic.ts) — the union of cursor-on-line ∪ hover ∪ selected/active ∪ panel-open ∪ crop-active — that every cross-cutting decision now reads instead of a fresh per-site filterPanel || classPanel || submenu || cropEditor chain. main.ts replaces those scattered chains in the click / overlay-mutation / Escape / hover-leave / float-region paths with anyPanelOpen() (filter/class/sub-menu subset, excludes crop) and anySurfaceOpen() (the full plugin-surface union the Esc / overlay-dismiss / hover-leave / float-region paths consult), both routed through isEngaged. The hover/cursor/selection inputs are gathered live by callers; only the union is pure. Routing the same cross-cutting decision through one predicate keeps the reveal pin (Bug 86), the floating-region dismiss and the overlay/Esc dismiss consistent. Realizes the pre-recorded AD12; implementation-plan.md updated (gated) (toolbar-region-logic.ts, main.ts).
• Decision 32 — Raw-link reveal derives from Obsidian's own parse via one pure decision resolveLinkReveal (AB16b / D16 / D17). The headline of the rework, on the model side: the reveal/dismiss of an image's raw markdown link no longer reacts to the DOM or a parallel regex — it is DRIVEN top-down from Obsidian's own parse spans (is the cursor within the body ![](…) span? within the {…} span?) plus the mode/dismiss/engaged state. A single pure, unit-tested function resolveLinkReveal (live-preview-logic.ts) computes the whole-link outcome: the body shows as EITHER the native raw link OR the plugin's display-only stand-in (fake link), mutually exclusive by construction (D16 — cursor-in-body ⇒ native carries it and the stand-in hides, otherwise the stand-in carries it), and the trailing {…} shows/hides as one whole with the body (D17). It returns showStandIn / reserveStandIn / showAttr / suppressNative; the StateField turns this into a top-down line marker class the CSS keys on with plain parent→child selectors (no :has). This realizes the already-recorded AB16b/D16/D17/AD10 architecture (architecture.md unchanged — it predates this code); implementation-plan.md updated to the resolveLinkReveal-driven model (gated). No behaviour-defining doc above the plan changed.
• Change 51 — Embed detection is now Obsidian's own parse (syntaxTree), not a parallel regex (AD10). What: the new collectEmbeds in live-preview.ts enumerates embeds from the editor syntaxTree — a markdown embed starts at an image-marker node, a wikilink embed at a formatting-embed node — and the regex now only parses the span the parse located (the ![…](…)/![[…]] body plus the trailing plain-text {…}), with placement (standalone / bare-block / inline) read from the surrounding line. Why: a code-block ![](…) carries neither marker node (only hmd-codeblock), so it is excluded by construction — no separate code-block check — while F43 "render in code blocks" re-includes those via a regex pass. ensureSyntaxTree forces the parse to document end so lower-line embeds appear in the same build; a parse timeout/failure fails open to a full regex scan so embeds never vanish; and a guarded updateListener dispatches a refresh when the background parser advances so newly-parsed regions render.
• Change 50 — The inline render-fork is dissolved — inline embeds use the same uniform chrome as standalone/block. What: live-preview.ts drops the separate inlineEmbeds path (and the EMBED_LINE/inlineEmbeds imports); the early if (this.mode === "inline") minimal-render branch in EmbedWidget.toDOM is gone, so the chrome assembly is one uniform path for all three modes — only the wrapper class differs (for CSS placement) and lie-float just routes a too-small icon to the floating toolbar. Inline embeds now flow through the same collectEmbeds + per-embed reveal decision as standalone and bare, getting the same lie-show stand-in and {…} mark. Why: a single mechanism for every embed (R0/uniform-rendering, AD3) — removes the divergent inline minimal path that caused the inline reveal glitches.
• Change 49 — A reveal flip updates the block stand-in in place — no image/caption rebuild (kills caption flicker + the resize-affordance 1c regression). What: the EmbedWidget signature now splits into a structuralSig() (everything that needs a full rebuild) and the reveal state (showStandIn/reserveStandIn); a new updateDOM mutates only the hosted block stand-in's reserve-triad class in place (via blockFakeClass(), keyed off a data-lie-struct stamp on the wrapper) and returns false only on a real structural change. Why: previously a reveal flip made eq() false, so CodeMirror destroyed and rebuilt the whole widget — re-creating the image and its async caption, which produced a visible caption flicker and the resize-affordance 1c regression. Now the image and caption DOM survive a reveal change untouched.
• Change 48 — Reveal/suppress/reserve CSS reworked onto plugin-owned markers (AB16b) — no :has, no !important. What: the reveal block in styles.css is rebuilt to key on plain parent→child selectors over line/element markers the StateField owns and sets in the same transaction, instead of guessing the cursor line from Obsidian's DOM: per-element .lie-fake-link.lie-show / .lie-attr.lie-show (driven by the AB16b decision), a new RESERVE TRIAD for the bare/raw-link BLOCK stand-in — .lie-fake-link-block.lie-reserve (an empty placeholder holding a line of height above the image, display:block; visibility:hidden) and .lie-fake-link-block.lie-reserve.lie-show (paints it, flipping only visibility — so reserve↔reveal is a pure paint swap with no reflow; standalone/inline embeds keep the legacy collapse/show), and a line-level .lie-suppress-native that hides Obsidian's revealed body tokens (.cm-line.lie-suppress-native > .cm-image, > .cm-url:not(.lie-attr)) on a <>-dismissed line. The old mode classes (lie-rev-auto/lie-rev-always), the .cm-active + .cm-line:has(> .cm-formatting-*) reveal-slave selectors and the .lie-dismissed … display:none !important overrides are removed. Why: mutual exclusion of native-vs-stand-in is now by construction (D16 — the StateField withholds lie-show when the cursor is in the body, so Obsidian's native source carries it) and the trailing {…} shows/hides as one whole with the body (D17), so the brittle :has/.cm-active/!important machinery is gone and the dismiss can suppress even Obsidian's own active-line reveal (Bug 65's render half).
• Change 47 — reduceReveal generalized to three modes with a per-mode dismiss auto-clear. What + why: the transient <> dismiss/hover state reducer reduceReveal (live-preview-logic.ts) now takes the three-valued mode (native/auto/always) instead of the old alwaysShow boolean, and its auto-clear ("a dismiss clears when you LEAVE the image, not sticky across visits") follows the mode: NATIVE clears off the active/cursor line; AUTO additionally keeps a dismiss on the hovered line; ALWAYS never auto-clears (a dismiss persists until toggled/reload). A line freshly toggled in the same transaction stays exempt, so a <> dismiss takes effect in its own transaction and only auto-clears on a later leave/cursor-move — holding for both the mouse and the :focus-within/keyboard path. The broader AD12 engaged pin is a display freeze layered on top and only ever keeps a dismiss longer.
• Change 46 — Migrate the legacy boolean alwaysShowLink to the three-state defaultRevealState on load. What: loadSettings now migrates a stored alwaysShowLink boolean to the new defaultRevealState reveal mode that supersedes it — true → "always", false → "auto" (the old false meant "auto / on-hover") — only when the new key was not already persisted, then deletes the stale key; the plugin also now reads () => this.settings.defaultRevealState (in place of alwaysShowLink) when constructing the LP extension. Why: the reveal-source setting was reworked into a single combined dropdown (the Feature 33 UI lives in settings.ts); Object.assign over DEFAULT_SETTINGS won't translate the old boolean, so existing vaults need this explicit one-time mapping to keep their prior preference (main.ts).
• Change 45 — Settings that affect the LP render now force an immediate live re-render of the open editor. What + why: refreshLivePreviewDecorations() is made public on main.ts and forces a re-render of every open MarkdownView's LP decorations; it is now called from the settings tab after toggling reveal mode, captions, the new code-block render and the toolbar — and from each editing-surface onClose (filter/class/size/crop). Otherwise the open editor kept its old decorations until an unrelated edit, so a setting change (e.g. switching reveal to auto/always) looked like it "didn't take" because hover still behaved by the old decorations until the user typed; forcing the refresh applies the new render immediately.
• Change 44 — Redraw the inline layout icon to signal the vertical alignment axis. What: The two text strokes of the inline layout icon moved from the box foot (y=16.5) to the box vertical middle (y=12) in both the source artwork layout-icons.ts and the synced preview layout-icons-preview.html; the file's leading doc-comment was rewritten to explain the new framing. Why: With inline redefined as the lone VERTICAL-alignment axis (vertical-align of an in-line image), the icon now depicts the text line meeting the box at its vertical middle rather than sitting at its foot — visually distinguishing it from the five HORIZONTAL block/float states. Artwork-only change; the layout-state behaviour and mapping are untouched.

• Change 43 — A "Syntax & info" settings card: full {…}-syntax help, editor-style highlighting + a self-info link. A new card at the bottom of the settings documents how edits are stored. Its heading carries a native book-open Lucide icon on the left (D12-coherent — chosen over an emoji, which would also clash with our no-innerHTML/no-runtime-<style> constraints) and a "More info" CTA on the right — prefixed with Obsidian's external-link glyph — that opens this plugin's own page in Obsidian's community-plugin browser (obsidian://show-plugin?id=<manifest.id>, via the existing openPluginStore helper). Above the framed list (in normal body text, outside .setting-items): an intro line, then a copyable code sample in a native "example" callout (data-callout="example", via a shared renderCallout helper), then a prose hint that width+height together distort and that one dimension + aspect-ratio is the responsive-friendly choice (keyword mentions as inline <code>). The sample is highlighted across the whole ![…](…){…} link, not just the {…} block — link text/target in accent (.lie-syntax-link), punctuation muted (.lie-syntax-punct), each {…} attribute its own underlined token (.lie-syntax-token), tokenized by renderEmbedPrefix/renderExampleLine and built from spans since the card is outside CodeMirror (no .cm-*); it wraps instead of scrolling and is selectable to copy. The framed list itself is one native Setting row per keyword (the literal token with its possible values as the row name in <code>, the explanation as its desc, also selectable), grouped size → layout → orientation/crop → appearance → classes/raw/marker and covering the full writer vocabulary (§2.3): width=/height= (CSS length — a bare number is px, a unit also works), aspect-ratio=, align= (float vs block, each value explained), rotate=, flip= (each value explained), transform= (crop placement), filter=, .classname, style= (raw-CSS escape) and the optional .lie claim marker (the plugin no longer emits it, but CLAIM_SELECTOR still honours it so a user can force export / standalone-runtime pickup). All localized (DE/EN) with the literal tokens kept as code; styling stays in styles.css (no runtime style injection). Documented in implementation-plan.md under the F20 Settings umbrella (no requirements change — in-app help) (settings.ts, styles.css, src/i18n/).

• Feature 43 — New "Render images in code blocks" setting (Live Preview only, default off) — the lone AD10 override. What + why: added the renderImagesInCodeBlocks boolean setting (settings.ts, default off, threaded through to the LP extension in main.ts) with DE/EN strings. Off (the default) defers to AD10 — an image embed inside a fenced or inline code block stays literal code, matching Obsidian's parse, which treats a code-block embed as not-an-embed. On, the plugin renders it as an image anyway; this is the single documented override of AD10's "defer to Obsidian's parse". Surfaced under F20 in the General settings card.

• Feature 33 — Reveal-source setting is now ONE combined dropdown (native / auto / always) replacing the boolean alwaysShowLink. What: the old Always show the link source toggle (auto/always) becomes a single three-state dropdown — native (source on the active/cursor line only, the new default), auto (additionally on hover), always (everywhere) — surfaced in the General settings card with new DE/EN strings. The setting key changes from the boolean alwaysShowLink to defaultRevealState: RevealMode (settings.ts), with the new RevealMode = "native" | "auto" | "always" type added in live-preview-logic.ts; a loadSettings migration in main.ts translates a stored legacy alwaysShowLink (see Change 46). Why: Feature 33 asked to merge the scattered reveal toggles into one ordered choice. The no-reflow-on-reveal behaviour that Feature 33's spec splits into separate "Auto — Höhe sichtbar" vs "Auto" options is instead baked into the model for ALL modes via resolveLinkReveal's reserveStandIn (the stand-in holds the source box as an invisible placeholder unless the native carries it), so the image never jumps on reveal without needing a fourth option; the explicit "Hidden" option is the per-image <> dismiss, not a mode.
• Bug 117 — Floating bar for inline-icon / short-block images dismissed before it could be reached. Cause: the floating hover bar sits ABOVE the image with an 8px gap, but dismiss was an immediate-on-leave check in the delegated mouseover handler — moving the pointer image→bar crossed that gap, left the image, and dismissed the bar before it could be reached (acute for a tiny inline icon). Fix: the floating bar now rides the SAME bindRegionHover active region (D6) as the panels: a new bindFloatRegion binds the image wrapper + the .lie-toolbar-floating element as ONE region with the shared 160ms travel-grace, and dismiss fires only when the WHOLE region is left (and no panel/palette/surface governs it). The region cleanup (floatRegionCleanup) is unbound on open of a new image and on dismissToolbar (main.ts).
• Bug 116 — Floating toolbar flung far above a cropped/scaled image, detached and unreachable. Cause: positionAbove anchored the floating bar to img.getBoundingClientRect(), but a cropped/scaled image's own <img> rect overflows its visible frame (the .lie-image-area clips it), so the bar was placed above the raw (oversized) image top rather than the box the user sees — landing far above the visible image, detached and unhoverable. Fix: anchor to the visible box instead — (img.closest(".lie-image-area") ?? img).getBoundingClientRect() — so the bar sits a gap above what the user actually sees and hovers; falls back to img when no image-area is present (toolbar.ts).
• Bug 115 — An INLINE embed's {…} attribute list is now highlighted as link syntax in Source mode. Cause: in Source mode lineDecorations matched only the whole-line EMBED_LINE, so a {…} attached to an embed that had text or a trailing character around it on the line (an inline embed, not a standalone) was skipped and left un-highlighted. Fix: Source mode now scans the whole line with a new EMBED_WITH_ATTR regex (every embed that carries a required {…}, standalone AND inline) and emits the brace/url-string marks for each match (live-preview-logic.ts). The LP-branch (standalone widget) path is unchanged.
• Bug 114 — A bare embed's markdown/wikilink source never revealed on hover in Live Preview. Cause: a bare embed (![](…) / ![[…]] with no {…}) is block-promoted by Obsidian into a cm-line-less .cm-content child, which swallows an inline stand-in decoration — so the plugin's block:true widget rendered the image but carried no .lie-fake-link at all, leaving the F8 reveal nothing to show. Fix: the block-mode EmbedWidget now hosts the stand-in raw link inside its own DOM (a .lie-fake-link-block span built in toDOM, the source for it carried by the per-embed reveal.showStandIn/reserveStandIn), so cursor / hover / always-mode / the <> dismiss all drive the bare source exactly like a standalone embed. Confirmed by the new bare/wikilink rows of verify-reveal-matrix.mjs (real-pointer hover, cursor-on-line, dismiss, engaged-pin).
• Bug 86 — Link-reveal flipped during a crop / panel interaction, making the image jump. Cause: the raw-link reveal followed the live cursor/hover state, so while a crop (or filter/class/size panel) held an image the reveal could toggle as the pointer or caret moved on/off the image's line — reflowing the line and shifting the image mid-interaction. Fix: an engagement-keyed reveal PIN (AD12 / AB16b) plus a no-reflow stand-in. (Model) resolveLinkReveal treats engaged as a natural-show with suppressNative:false so the link stays visible and does not flip (live-preview-logic.ts); a new engagedImagePos() in main.ts resolves the exact document position of the engaged embed (via posAtDOM on the active image's .lie-wrapper, gated on anySurfaceOpen() + a connected active image, read through the pure isEngaged predicate — see Decision 33) and is passed into createLivePreviewExtension, so the StateField pins THAT exact embed precisely — a standalone occupies its whole line, an inline embed only its own span, so a sibling inline embed on the same line is not over-pinned, and it returns null when no surface is open or the active image can't be resolved. (Render) build reads getEngagedPos() and passes engaged into the per-embed decision so the reveal is pinned and cannot flip (live-preview.ts), while the block stand-in sits on its own reserved placeholder line above the image — a RESERVE TRIAD in styles.css (.lie-fake-link-block.lie-reserve is display:block; visibility:hidden, .lie-reserve.lie-show flips only visibility) — so reveal↔hide is a pure paint swap with no reflow and the image never jumps. The auto-clear is also engaged-aware: since ENGAGED ⊇ cursor∪hover it only ever keeps a dismiss longer, never clears one earlier. The engaged-pin case is covered by the bare engaged-pin off-line row of verify-reveal-matrix.mjs.
• Bug 65 — A <> dismiss now suppresses Obsidian's native raw link even on the cursor line. Cause: on the active/cursor line Obsidian reveals its OWN editable raw-link tokens (the real ![…](…) document text); the old <> dismiss hid only the plugin's overlay (.lie-fake-link + .lie-attr via .lie-dismissed), so it lost the fight with the native reveal and the front of the link stayed visible. Fix: AB16b makes the StateField the single authority. (Model) resolveLinkReveal returns suppressNative:true on a dismiss ({ showStandIn:false, reserveStandIn:false, showAttr:false, suppressNative:true }) so the plugin overrides even where Obsidian would natively reveal, and collapses the reserve box (the user explicitly cleared the source, so no placeholder gap is left) (live-preview-logic.ts). (Render) the StateField stamps the line marker lie-suppress-native, and CSS hides Obsidian's revealed body tokens scoped to that line (.cm-line.lie-suppress-native > .cm-image, > .cm-url:not(.lie-attr)) using plain parent→child + :not, no :has, no !important — the native text stays in the document and editable, only the on-screen reveal is suppressed (Lesson 11/12). CDP-verified (styles.css).

[0.6.10] - 2026-06-25

Crop-overflow rework — Obsidian 1.12.7 raised the specificity of its block-widget contain:paint, which silently defeated the crop editor's contain:none lift, so the croppable surround and the resize frame/corner were clipped at the cut window. Replaced the brittle lift with a body-portal model that stops fighting containment. This release also lands a black-box / optical + functional CDP test layer (Change 41) — pixel + pointer-hit-test checks that observe the painted result instead of reading CSS properties, with a one-command runner — plus an honesty pass removing false-greens and un-runnable false-reds from the existing checks (Change 42); exactly the layer whose absence let this regression ship unseen. It also makes the hover resize marker mirror Obsidian's native handle exactly and stops it being clipped on block-layout images (Bug 112) — a padding-only reserve, no contain:none — with new optical coverage (block marker not clipped, caption anchor). Verified live via CDP on 1.12.7 (the full verify-all suite now 15/15, incl. verify-resize-affordance 9/9 with the block/caption marker checks and verify-crop 19/19) plus npm test green (283), npm run build / lint / lint:css clean and lint:obsidian 0 errors.

• Change 42 — Test honesty pass: false-greens and un-runnable false-reds removed from the CDP checks. A green on a broken feature and an always-red check that can't drive what it asserts both make the suite lie, so the existing checks were corrected against the current behaviour: verify-render-gaps now expects the icon size preset to persist a line-height height=, not .lie-inline (Change 38 decoupled size from layout — icon is a height, inline is a separate layout state); verify-write-path drives the current applyLayout("float-left") / applyLayout("inline") instead of the removed applyAlignment / toggleInline (Decision 30); verify-crop drops the inline transform-origin read (empty by design — the centre pivot lives in CSS, so the read was always-red regardless of behaviour; the pivot is covered behaviourally by preview == committed); and verify-submodal-region is re-driven by a real CDP pointer (synthetic MouseEvents can't drive the binder's real :hover / pointer-events path — they false-RED the leave checks AND false-GREEN the "still active" ones, since the region never actually moved). (test-plan.md records the visibility / "never assume" rules behind this — gated, user-approved.) (tests/cdp/).
• Change 41 — Black-box / optical + functional CDP test layer (observe the painted result, not CSS). A regression layer that verifies the running plugin the way a user sees it — sampling rendered pixels (region screenshots decoded to RGBA) and pointer hit-tests (elementFromPoint, a REAL CDP :hover via Input.dispatchMouseEvent) instead of reading CSS properties, so checks survive CSS refactors and Obsidian updates (the gap that let Bug 111 ship unseen). It adds a reusable client tests/cdp/_optical.mjs (evaluate / real hover / screenshot → RGBA PNG decode); verify-optical-render (geometry — rotate/flip footprint, column cap, caption below+within-width, block/float/inline, both views), verify-optical-pixels (rotate/flip content + grayscale, sampled), verify-optical-chrome (toolbar / sub-menu / handle visibility, not mere presence); verify-functional drives the size modal as a user and reads the rendered width and the persisted {…} source back; and a one-command runner verify-all.mjs that runs every verify-*.mjs and sums pass/fail ("test a new Obsidian version" in one go). (test-plan.md extended to the optical / functional / visibility model — gated, user-approved.) (tests/cdp/, docs/development/test-plan.md).
• Bug 112 — On a block-layout image the native resize marker was clipped by the block widget's contain:paint. Cause: the plugin's resize marker deliberately deviated from Obsidian's native handle — it sat centred on the corner tip, half outside the image — so on a block:true widget (tall-float / bare embed), which Obsidian paint-contains (.cm-content > [contenteditable="false"] { contain:paint }), the outside half was cut off where the image edge met the widget edge (the same root as Bug 111, but for the always-on hover marker, not the crop chrome). Fix: make the marker mirror Obsidian's native handle exactly (D4 — now a fixed requirement that also rules out "just nudge it inside") — the dot sits flush in the corner with only its 2px accent outline bleeding out, matching app.css div.image-embed .image-resize-corner (box -3px, marker +3px) — and reserve that 2px inside the clip with a padding-only reserve on .lie-wrapper-block (padding-inline-end + padding-bottom). The block widget is over-constrained, so a compensating negative margin is dropped by the engine (verified — even !important resolves to 0); the padding is absorbed by shrinking the content, so a full-width block image gives up 2px (0.3 %, D3-safe, no overflow; standalone images carry no containment, untouched). No contain:none override — that brittle pattern is exactly what Bug 111 retired; a body portal for the marker stays the documented alternative if the 2px ever matters. Regression-guarded by the optical layer: verify-resize-affordance now checks the native marker on a block image is fully painted (both the right and bottom 2px bleeds survive the clip — proven by removing the reserve → red) and that with a caption the handle anchors to the image corner, not the caption's bottom (closes the Bug 99 / Bug 107 coverage gap); verify-optical-chrome's marker check was re-tuned to the native geometry. requirements.md (D4), test-plan.md and implementation-plan.md (AD3) updated (gated; user-approved) (styles.css, tests/cdp/).
• Bug 111 — Obsidian 1.12.7's contain:paint specificity bump broke crop overflow; the resize frame / corner was clipped. Cause: the in-place crop editor lifted the Live-Preview block-widget host's contain:paint with a doubled-class contain:none !important so the dim ghost + handle chrome could overflow the cut window; Obsidian 1.12.7 raised the host rule's specificity, so the lift lost and the host re-clipped everything to the footprint — the croppable surround vanished and the resize frame/corner were cut off. Fix: stop fighting containment (Variante B) — the host's contain:paint is now honoured and the in-host result image stays clipped as the footprint; everything that must extend past it (the dim ghost showing the croppable surround + the handle chrome) moves to a body portal anchored to the in-host frame's rect (re-anchored on scroll/resize), with a clip-path hole over the result so the in-host result shows through. The hole is path(evenodd, "<outer rect> <hole rect>") — two contours making a true frame; a single polygon() joins the rects with diagonal edges that slice triangular artifacts into the surround (caught + regression-guarded by the new optical pixel check). Pan works on both surfaces (the in-host result through the hole, the portal ghost outside); teardown removes the portal and its scroll/resize listeners on every exit path. architecture.md (AB12), implementation-plan.md and test-plan.md were updated to the portal model (gated; user-approved) (crop-editor.ts, styles.css).

[0.6.9] - 2026-06-16

Re-release after the community-plugin review failed v0.6.8: fixes the one review-failing error and clears the largest warning class. Verified npm test green (283), npm run build / lint clean, lint:obsidian 0 errors (17 documented warnings, down from 101).

• Change 40 — document → activeDocument across the in-Obsidian plugin source — the first part of Feature 39 (popout support). This is one half of the still-open Feature 39; it lands the activeDocument conversion only (the window → activeWindow half and popout-reachability verification remain open under Feature 39). The review bot flags obsidianmd/prefer-active-doc as a WARNING on every bare document reference, and the rating counts per occurrence — ~95 hits dragged it down. All global document references in the in-Obsidian plugin source were converted to Obsidian's window-aware activeDocument (popout-correct; identical to document when there is no popout, so no behaviour change in the common case). The standalone runtime (runtime.ts) keeps raw document — that off-Obsidian bundle imports no obsidian, so the activeDocument global does not exist there (same exception class as its raw instanceof); its 6 flags stay documented warnings. Done by replacing only the exact eslint-flagged code positions (not a blind text sweep, so comments/strings naming "the document" are untouched). lint:obsidian's prefer-active-doc rule is now warn (was deliberately off) so a regression resurfaces. This is the activeDocument portion only; the remaining Feature 39 work (window → activeWindow, and confirming popout reachability) stays open (eslint.obsidian.config.mjs, plugin source).
• Bug 110 — The standalone runtime's caption renderer assigned innerHTML, failing the community review. Cause: v0.6.8's new off-Obsidian caption renderer (Feature 41) set caption.innerHTML = renderInlineMarkdown(...), which the review bot rejects (Unsafe assignment to innerHTML) — the single error that failed the v0.6.8 review. Fix: renderInlineMarkdown still returns its already-escaped, href-sanitised HTML string (kept a pure string→string unit, its 6 tests unchanged), but the runtime now parses it via DOMParser and grafts the nodes in with replaceChildren instead of writing innerHTML — the parse runs in a detached document (no script execution) so, with the upstream escaping, the result is inert (runtime.ts, runtime-markdown.ts).

[0.6.8] - 2026-06-13

Captions, crop and resizing fixes — the standalone runtime (lie-runtime.js) now renders image captions off-Obsidian (with inline Markdown); the resize handle anchors to the image corner even with a caption; resizing on hover no longer breaks undo; and the link-source reveal stays put during a crop. Verified npm test green (283), npm run build / lint / lint:css clean.

• Decision 31 — Off-Obsidian the runtime carries its OWN minimal inline-Markdown caption renderer (an explicit AD9 exception). AD9 ("reuse the platform, never a parallel reimplementation") binds only where the platform actually provides the capability. On a foreign page the standalone runtime has no Obsidian and no MarkdownRenderer, so it may carry a small runtime-only inline renderer (bold / italic / code / link) for captions — that is not a parallel reimplementation of a platform capability (there is none to reuse here), and the plugin's caption path keeps using MarkdownRenderer unchanged. Full fidelity stays bounded by the lossy alt attribute (e.g. python-markdown strips code-span backticks before the runtime sees the alt). Recorded as the AD9 "runtime exception" in architecture.md.
• Feature 41 — The standalone runtime now renders image captions off-Obsidian. Previously lie-runtime.js built the 3-layer image but never the caption, so captions worked only inside Obsidian (the reported "standalone captions don't work"). The runtime now reads a claimed image's alt, derives the caption text with the plugin's own captionFromAlt (native size token stripped) and wraps the image in a .lie-has-caption shrink-wrap host — the off-Obsidian stand-in for the plugin's .lie-box — sized to the image width by CSS alone (D9, no JS width-sync). The caption shell + host builder live in a new shared, Obsidian-free caption-dom.ts reused by both the plugin and the runtime; the text is rendered with the runtime's own inline-Markdown renderer (Decision 31). A claim trigger is required, so the width-only caption examples in the example vault carry a .lie marker (caption-dom.ts, runtime.ts, runtime-markdown.ts, render-core.ts).
• Bug 109 — While cropping, the link-source reveal flipped as the mouse left the image. Cause: the crop chrome (rotate handle etc.) extends beyond the image / crop-frame, so dragging the pointer onto it leaves the .cm-line, and the auto reveal is keyed on .cm-line:hover — so the source link popped in and out mid-crop. Fix: suppress the hover reveal on a cropping line (.cm-line:not(:has(.lie-cropping))); during a crop the source follows ONLY the stable signals (.cm-active, always-mode, the <> dismiss), so an off source (auto-mode resting state or dismissed) stays off and an always-shown one stays shown, regardless of where the crop drags the mouse (styles.css).
• Bug 108 — Resizing via the hover handle without first clicking the image sent cmd+Z to the document top. Cause: the resize handle is reachable on hover, so the caret may sit at offset 0; the resize write did not pass the D11 cursor, so undo restored the offset-0 selection and scrolled to the document top. Fix: the resize write now passes the resized image's line as the cursor (D11), seeding the change transaction's startSelection so cmd+Z stays on the image (live-preview.ts).
• Bug 107 — With a caption the resize handle sat at the caption's bottom-right, not the image corner. Cause: the handle is position:absolute; bottom:0 against .lie-box, which a caption (a sibling flex item, D9) extends below the image — so bottom:0 anchored to the caption's bottom. Fix: wrap the image and the handle in a non-clipping .lie-image-host that shrink-wraps the image only, so bottom:0 anchors to the image corner; the box stays the caption shrink-host (D9), and the corner marker keeps its centred-on-the-corner position because the host has no overflow:hidden (unlike the image-area, RENDER_CSS) (live-preview.ts, styles.css).

[0.6.7] - 2026-06-12

Inline and in-list/callout image embeds now reveal their link source exactly like a standalone image — fixing the long-standing "the link shows without its {…}" / "stray {…} with no link" glitches when the cursor is on the line. The inline decoration path was rebuilt onto the standalone reveal machinery and the reveal heuristic was narrowed; verified live via CDP across standalone / inline / list / markdown-link / native-size embeds, npm test green (277), npm run build / lint clean.

• Bug 106 — List/callout image showed the {…} PERMANENTLY as stray text, the link never revealed. Cause: the reveal heuristic .cm-line:has(> .cm-formatting) (hide the fake link when Obsidian natively reveals the source; show the {…}) matched ANY direct cm-formatting child — a list bullet (cm-formatting-list) or a callout/quote marker — so on a list/callout image line it fired in EVERY state: the fake link was always hidden and the {…} always shown (a stray attr with no link). Fix: narrow the selector to the embed's OWN link/image tokens (cm-formatting-link for a wikilink, cm-formatting-image + cm-formatting-link-string for a markdown link), which a structural list/quote marker never carries (styles.css).
• Bug 100 — Inline (mid-text) image: the LINK wasn't shown, only the {…}, and the {…} had no syntax highlighting. Cause: the inline path replaced the whole ![…]{…} with a cursor-skipped Decoration.replace and gave the {…} none of the standalone path's fake-link/.lie-attr treatment; with the cursor on the line the replace was dropped entirely, so the native image (uniformly CSS-suppressed) vanished and the bare {…} was left as stray text with no link. Fix: the inline path now mirrors the standalone path (architecture AB16 — "inline embeds get the same widget; only chrome placement differs"): a display-only fake link paints the source for the reveal-for-looking, the {…} is a marked + url-string-highlighted native text, and the plugin draws its own inline image widget; no more cursor-skipped replace (live-preview.ts).
• Bug 96 — Link shown DOUBLED in Live Preview (Doppellink). Cause: the reported "standalone double" was an INLINE image with the cursor on its link — the cursor-skipped inline replace (Bug 100) fell back to Obsidian's native (image-suppressed) embed with no coordinated fake link, so the source could surface alongside the painted artifacts. Fix: resolved by the same inline rework — the fake link now yields to a GENUINE native reveal via the narrowed :has selector (Bug 106), so the link is never shown twice (live-preview.ts, styles.css).

[0.6.6] - 2026-06-10

A six-state layout rework (block/float/inline), a native-size wikilink fix and several smaller fixes — verified npm test green (277), npm run build / lint / lint:css / lint:obsidian clean.

• Decision 30 — Layout is ONE flat six-state choice (block-left/center/right · float-left/right · inline), not two conflated axes; serialized back-compat + HTML-faithful. The old model overloaded align=left|right to mean float and kept inline as a separate boolean, so "block-left/right" (aligned, no wrap) was impossible and the toolbar couldn't show which state was active. The rework makes the six valid states a single layout field (the impossible combinations — float-center, inline-with-position — simply don't exist), surfaced as six radio buttons (exactly one active). On disk: float stays the genuinely HTML-faithful align=left|right (a browser floats <img align=left>), block uses the align=block-left|block-center|block-right keys, inline the .lie-inline class; legacy align=center and .lie-left|right|center classes still READ (migrate on next save). Chosen over the symmetric align=left=block scheme because that would have silently reinterpreted every existing float note AND dropped the float HTML-fidelity. The earlier "click-cycles, hover-opens-submenu" idea was dropped for the simpler flat six-button set (no invalid combinations, nothing to gray out). Icons are hand-drawn (Lucide has none for image-to-text layout; a gallery-vertical family — frame + box / box + side-lines / box-in-a-line), since icon artwork is design-patent/copyright territory only for a specific drawing, not the generic concept.
• Change 39 — The editing-toolbar submenu now inserts in the MIDDLE of the bar, not the left edge. The original slot (right after a leading undo/redo run) put our entry at the far left, where its expand-on-click row reaches off the right window edge on a wide bar ("ragt aus dem Fenster", per the user). insertIndexFor now returns floor(list.length / 2). We still only set the INITIAL slot — a re-add never reorders an existing, correctly-formed entry, so a user's manual placement is kept.
• Change 38 — height= serializes as a bare key (round-trip); size is decoupled from layout. The parser already read a bare height= (applyKey), but serializeTransform always normalized it to the style="height:…" escape, so {… height=1.4em} didn't round-trip. It now emits height=N as a bare key for every unit (a pure px value drops the unit, mirroring width=N). Separately, the size sub-menu's SizeState/presets no longer carry an inline flag — the icon preset sets only a line-height height; the inline rendering is the inline layout state, chosen independently.
• Feature 40 — The six-state Layout group: radio buttons + active highlight + custom icons + hover animations. The Layout toolbar group is six buttons (block-left/center/right, float-left/right, inline), exactly one highlighted to match the image's current state (computed per show in toolbarItemsForImage; the .lie-toolbar-btn.is-active accent rule — the first toolbar active-state, mirroring .lie-class-item.is-active). Each carries a custom addIcon icon (layout-icons.ts) and a fitting hover micro-animation (block = directional nudge, float = a sideways "wrap" drift, inline = a bob into the line). The same six states drive the command palette, the editing-toolbar submenu and the keyboard commands. CSS: float floats the host, block uses text-align on a full-width host (block-left/right are new), inline keeps the direct vertical-align rule; the standalone runtime mirrors all of it with direct selectors.
• Bug 94 — A native wikilink/markdown size was ignored when rendering. Every render path built the transform from the {…} block alone (parseAltText(params)), so a raw ![[img|160]] or ![[img|160 "Caption"]] — size in the link, not the block — rendered at the default column size (the caption already parsed, only the size was dropped). The native size token (already extracted by link-format's splitTail) is now folded into the transform's width/height on read in BOTH the reading-view post-processor (processBlock / reconcileFromSource) and the live-preview widget, via the new pure applyNativeSize (the explicit {…} block always wins — same precedence as the write path). EVERY pipe variant is covered: bare 300, 300x200, the auto-axis forms, a size with a caption in either order, the legacy caption|300 suffix, and the markdown ![alt|300](p) form — in both link forms. Live preview reads the raw source; reading view uses a priority chain (raw source alias → img.alt token → the width/height attributes Obsidian sets for a pure native size, where it strips the size from the alt) so no variant slips through (foldNativeSize). On an actual EDIT the size is normalized into the block and stripped from the link head (writeTransform rebuilds the embed via buildEmbed; modifyTransform/the size, crop and filter panels read the current size through the shared locationTransform), keeping the design invariant that size lives in the block, never the pipe (F6/T2). A plain caption-only embed is untouched.
• Bug 79 — The editing-toolbar submenu could keep stale pre-rework layout entries. The load-time self-heal that migrates our submenu to the current command set (here: the six new layout entries, replacing the old class-left|center|right / toggle-inline) shared one await chain with the snippet scan in onLayoutReady — await this.refreshSnippets(); await ensureEditingToolbarButtons(…). A refreshSnippets() throw (a vault read error) skipped the migration entirely, freezing the submenu at its previous shape until the next clean reload (observed live). The two are now independent: the scan is wrapped in its own try/catch so it can never block the migration.
• Bug 78 — A block-aligned image (align=block-* / legacy align=center) broke its source line onto two lines in live preview. The block layout rules styled both the plugin's .lie-wrapper widget AND the native .image-embed with display:block; width:100%. In reading view .image-embed IS the image container, so that is correct; but in live preview it is the SUPPRESSED native embed sitting inline in the editable source line, and forcing it full-width block pushed the {…} block onto a second visual line below the ![](…) link, even with ample width. The .image-embed half of the block rules is now reading-view-scoped (.markdown-reading-view / .markdown-preview-view); in live preview the standalone .lie-wrapper widget alone carries the block layout (it already centers the image), and the native embed stays inline. Float rules are untouched.
• Bug 77 — Undo after a toolbar edit jumped the view to the document top. Plugin edits were written without a selection (old D11 — "never move the cursor"), so the edit transaction's startSelection stayed at offset 0; CM6 restores that selection — with scrollIntoView — on undo, hurling the view to the top of the note. D11 is revised: a single-image edit now places the caret on its image's line (like Obsidian's own embeds), but only on an actual edit, never on hover. Mechanically, writeSource takes an optional cursor and, when given, dispatches a prior selection-only transaction (addToHistory: false) so the caret sits on the image before the change — making that position the change's startSelection, which undo then restores. (Setting the selection on the change transaction would not help: undo uses the selection from before the change.) Bulk writers — multi-select edits, link normalization — omit the cursor and are unchanged.
• Bug 76 — A non-native layout now CLAIMS, so block / center / inline render on a published page. CLAIM_SELECTOR claimed an image only on a runtime-only key; align=center (and now align=block-* / .lie-inline) have no faithful HTML rendering, so off-Obsidian they showed unstyled. They are added to the claim selector ([align="center"], [align^="block-"], .lie-inline + data- variants); align=left|right (float) stays unclaimed — a browser floats it natively. readTransform also now reads a height= attribute.

[0.6.3] - 2026-06-07

A one-bug fix to localization: the plugin now follows Obsidian's UI language faithfully, including when Obsidian is set to English. Verified: npm test green (272), npm run build succeeds.

• Bug 93 — The plugin ignored Obsidian's language whenever Obsidian was set to English. detectLocale() returned navigator.language (the OS/browser locale) whenever getLanguage() was "en", so on a German OS the toolbar/menus showed German even with Obsidian set to English — the opposite of F21 ("follow Obsidian's locale, no language setting of our own"). getLanguage() is documented to return the configured app language and default to "en", so the !== "en" branch made the plugin prefer the system locale over Obsidian's actual setting. Fixed: detectLocale() now mirrors getLanguage() verbatim (getLanguage() || navigator.language || "en"); the browser locale stays only as a last-ditch guard for an (per the API, impossible) empty return.

[0.6.2] - 2026-06-07

A follow-up to the v0.6.1 community-plugin-review compliance pass. The automated re-review of v0.6.1 (review-0.6.1.md) still failed — now on a single error: the standalone runtime's createElement("style") (runtime.ts:32). v0.6.1 had documented that file as a "false positive" and EXCLUDED it from the local recreation, but the real review bot scans it anyway, so the exclusion only hid the failure instead of resolving it. This release fixes it at the source — the runtime now injects its CSS via a constructable stylesheet (adoptedStyleSheets), not a <style> element — and makes the recreation bot-faithful: npm run lint:obsidian now scans all of src/ (incl. the runtime + dev-bridge non-plugin bundles) and mirrors the bot's severities, so the gate's error set equals the bot's and would catch this class before the bot does. Everything else the review flags is a documented warning/recommendation (Decision 26), none of which fails the review. Verified: npm run lint:obsidian 0 errors (11 documented warnings), npm run lint clean (T9), npm test green (272), npm run build succeeds.

• Decision 29 — The review recreation must SCAN everything the bot scans and MATCH its severities; non-plugin false-positives are fixed at the source or kept as documented warnings, never hidden by exclusion (supersedes the exclusion clause of Decision 25). Decision 25 excluded the standalone runtime (runtime.ts → lie-runtime.js) and the dev-only CDP bridge (dev-bridge.ts) from lint:obsidian as "false positives for the shipped plugin" — but the bot scans the whole repo, so the runtime's <style> kept failing the real review while our local gate stayed green (the drift that produced review-0.6.1.md). Corrected: eslint.obsidian.config.mjs no longer ignores those two files, and rules the bot reports as WARNINGS are set to warn locally (the recommended ruleset makes import/no-nodejs-modules and obsidianmd/prefer-instanceof hard errors; the bot — like our no-deprecated handling — shows them as warnings) so "0 errors locally" ⟺ "review won't fail". The genuine off-Obsidian/dev-only false positives are kept as documented warnings (runtime raw instanceof — no obsidian import, so no instanceOf() helper; dev-bridge net — tree-shaken from production) rather than silenced with inline disables that would break the shipped linter (T9 / Lesson 17). prefer-active-doc (the bot's document → activeDocument warnings) is large, popout-only churn with no review-failing impact and is deferred to Feature 39, kept off in the gate by deliberate decision (not suppression). → Lesson 18.
• Change 37 — The runtime injects CSS via adoptedStyleSheets instead of a <style> element; lint:obsidian made bot-faithful. runtime.ts's inject() no longer does document.createElement("style") + head.appendChild; it builds a new CSSStyleSheet(), replaceSync(css), and appends it to document.adoptedStyleSheets (same effect on a foreign page, no forbidden element — the Obsidian-review no-forbidden-elements error class is gone at the source). A module-level Set keeps the per-id dedupe the old getElementById guard provided. eslint.obsidian.config.mjs drops the runtime.ts/dev-bridge.ts ignores, downgrades import/no-nodejs-modules + obsidianmd/prefer-instanceof to warn (bot severities), declares the NodeJS + __LIE_DEV__ ambients so no-undef doesn't fire on recreation artifacts, and relaxes the no-console wrapper for the dev-only bridge. A new guard test (tests/unit/runtime-style-injection.test.ts) fails the moment a <style> element creeps back into the runtime (source-text assertion — the runtime is browser-only and can't load in jsdom). 272 tests green. No behaviour change for users (the runtime renders identically; the plugin proper was already <style>-free since Change 33).

[0.6.1] - 2026-06-06

An Obsidian community-plugin-review compliance pass. The automated review (eslint-plugin-obsidianmd + a CSS scan + a behaviour scan) failed on two error classes — inline static styles and a runtime-injected <style> element. Both are resolved with the smallest faithful change: static styles move to CSS, dynamic per-image values stay inline, and the plugin no longer hand-injects CSS. The review is now reproducible locally (npm run lint:obsidian / lint:css) without touching the shipped linter (T9). Independently of the review, this release also fixes a footprint bug when rotating explicitly-sized images (Bug 90). Verified: npm run lint:obsidian reports 0 errors (6 documented deprecation warnings), npm run lint stays clean, npm test is green (270), npm run build succeeds.

• Decision 28 — :has() is reserved for reacting to Obsidian's own DOM; self-set markers/classes ride the element they style (the outer). Reaffirms AD2/AD3: everything the user's {…} block carries — .class, style, the align marker — belongs on the outer .lie-image-area (the size/layout element), never the inner <img>. Adds the discipline that since the plugin BUILDS the outer (and the LP .lie-wrapper), every class/marker it sets is placed on the element that must react and selected DIRECTLY (.lie-image-area.lie-left) — :has() is used ONLY where the plugin must react to Obsidian's own, uncontrolled DOM (the source-reveal slaving .cm-line:has(> .cm-formatting)), never :has(img.lie-*) for our own classes. The code previously violated this (classes/markers on the <img>, alignment via :has(img.lie-*)) — the root of the decoration-clipping the reverted shadow/border workaround papered over — now FIXED in Change 36 / Bug 91 (markers on the outer; runtime :has-free; the plugin's host float :has reads the outer's marker, the tolerated CM case). Snippet contract: with the class on the outer, an example/vault snippet is authored as a plain .classname (+ .classname img for pixel effects like circle's object-fit), never the internal .lie-* structure. Recorded in architecture.md (AD3 note + AB20 + the F18 row).
• Decision 27 — Adopt GitHub build-provenance attestations for the release binaries. The release workflow now runs actions/attest-build-provenance on main.js / styles.css (permissions id-token: write + attestations: write), so a CI-cut release is cryptographically tied to this repo+commit (verify with gh attestation verify <file> -R <owner>/<repo>). Caveat recorded: attestation only happens inside GitHub Actions (OIDC/Sigstore) — there is no purely-local path, so the /release skill's local gh release create stays unattested; making CI the sole publisher is a separate future decision.
• Decision 26 — Keep the justified CSS-scan warnings and the 1.12.7-floor deprecations (documented, not removed). The remaining :has() warnings are unavoidable: alignment/float/reveal must style a flow-participant ANCESTOR (the embed / .cm-line) from a marker on a DESCENDANT (the outer box, or Obsidian's own .cm-formatting) — the only CSS-only mechanism (Bug 10/27; AD5 forbids reactive JS). The removable :has (inline, embed shrink-wrap) were dropped in Change 36; the runtime is :has-free. The !important warnings override Obsidian core / dynamically-gated rules (the crop contain:none beating app.css contain: paint !important, the tall-float cap, the dismissed-reveal) — audited in Change 36, the one removable defensive !important dropped. PluginSettingTab.display()
• ButtonComponent.setWarning() are deprecated since 1.13.0, but their replacements (getSettingDefinitions / setDestructive) are 1.13.0-only and minAppVersion is 1.12.7 (display() is the officially-sanctioned <1.13.0 fallback) — kept; no-deprecated set to warn in the recreation. All are warnings, not the errors that fail the review. The review's "Vault Enumeration" behaviour flag (vault.getFiles() in replace-picker) is likewise justified and left as-is: it backs the F26 Replace-image fuzzy picker (a FuzzySuggestModal needs the full image candidate set up front), is already filtered to image extensions, and has no narrower public API.
• Decision 25 — Recreate the review as a SEPARATE dev-only lint pass; the shipped linter is kept exactly as-is (T9). Added eslint-plugin-obsidianmd (recommended) behind npm run lint:obsidian (a new eslint.obsidian.config.mjs) and a stylelint npm run lint:css (stylelint.config.mjs) that reproduces the bot's :has/!important CSS scan — neither touches eslint.config.mjs / npm run lint. The standalone runtime (runtime.ts → lie-runtime.js) and the dev-only CDP bridge (dev-bridge.ts, tree-shaken via __LIE_DEV__) are excluded from the pass: off-Obsidian / in dev they legitimately need createElement("style") / raw instanceof / a net import, so the bot's flags on them are FALSE POSITIVES for the shipped plugin. They are excluded rather than silenced with inline eslint-disable comments because those would name obsidianmd rules the shipped linter doesn't know and would break npm run lint (→ Lesson 17).
• Change 36 — Everything the {…} block carries now rides the OUTER box, not the <img> (Decision 28 / Bug 91). render-core's applyClasses puts the user's .class list (incl. the built-in rounded/shadow/bordered/circle and any vault-snippet class) AND the lie-left/right/center / lie-tall / lie-inline markers on the OUTER .lie-image-area (one tracked set, data-lie-classes). This fixes the real defect behind the reverted shadow/border workaround: a decoration class now styles the un-clipped footprint box, so box-shadow / outset border are no longer clipped by the wrapper's overflow:hidden. The bundled snippet moves to the clean public contract — plain .rounded / .shadow / .bordered / .circle (no internal .lie-*), .circle img, img.circle { object-fit:cover } (the img.circle fallback for a bare/exported image), and theme-adaptive colours (var(--text-normal) border, a color-mix text-colour shadow). snippet-classify parses the new .NAME form (legacy img.NAME still accepted so an installed copy migrates; the auxiliary .circle img rule is not mis-counted), and resetLieState clears a reused img's legacy classes-on-img. Then the REMOVABLE :has went away: the inline :has(img.lie-inline) became the direct .lie-image-area.lie-inline, and the reading-view embed shrink-wrap :has(> .lie-image-area) became a .lie-embed class the post-processor sets on the host that holds the box. The standalone runtime is now :has-free — with the align marker on the outer (= the flow participant off Obsidian) it floats via a direct .lie-image-area.lie-left. The only :has left are the reveal (Obsidian's own .cm-line/.cm-formatting DOM) and the plugin's alignment-float family, now .host:has(.lie-image-area.lie-…) (reacting to the outer's marker) — the tolerated CM-context case where the plugin can't class the host it doesn't own. Exactly per Decision 28. BREAKING for snippets: a snippet (or installed bundled copy) targeting img.myclass no longer matches — Reset the bundled snippet in settings, and update hand-written vault snippets to .myclass / .myclass img. NOTE: the box-effect visuals + the inline/shrink-wrap layout are not headless-verifiable — verify in Obsidian.
• Change 35 — API-correctness / lint cleanups across the plugin (no behaviour change). instanceof HTMLElement → node.instanceOf(HTMLElement) (cross-window safe, main.ts); bare setTimeout → window.setTimeout (toolbar.ts); promise-returned-where-void-expected fixed with void / a threaded reject (main.ts export button, settings ConfirmModal — its onConfirm widened to () => void | Promise<void>); five redundant type assertions removed (main.ts, render-core align); loadData() laundered with as Partial<LieSettings>; the post-processor .bind(this) replaced with a typed arrow; the two English notices dropped their redundant "Live Image Editor:" prefix (sentence-case); localStorage["language"] replaced with Obsidian's getLanguage() (F21, prefer-get-language); @codemirror/state + @codemirror/view declared as devDependencies (Obsidian-provided externals — declared, not bundled, so T1 "no runtime deps" holds).
• Change 34 — Static inline styles → CSS classes (no-static-styles-assignment). The 3-layer centring/sizing literals move into RENDER_CSS (.lie-frame/.lie-frame > img, with a lie-crop-fit marker for the crop height:auto); crop-editor, anchored-submenu, toolbar and export static literals move to styles.css classes (.lie-crop-img, .lie-crop-frame-box, .lie-measuring, .lie-submenu-hidden, position/z-index on .lie-group-popup / .lie-toolbar-floating, .lie-export-path-input); the crop host contain:none override moves from inline-!important to a doubled-class styles.css rule (the same beat-app.css device as .lie-toolbar-btn.lie-toolbar-btn). Per-image DYNAMIC values (the crop placement transform, filter, computed widths/positions, the panel coordinates) stay inline — the rule only flags STATIC literals, so they need no change.
• Change 33 — Eliminate the plugin's runtime-injected <style> element (no-forbidden-elements). styles-injector no longer creates a <style>: the static RENDER_CSS layer rules and the built-in align/inline classes live in styles.css (Obsidian loads it), the preset-width vars are set on <body> via style.setProperty, and a disabled built-in class is a body.lie-cls-off-NAME marker the body:not(.lie-cls-off-NAME) rules react to — the same body-class device as applyTallFloatClass/applyButtonOutlines. RENDER_CSS stays the single source the standalone runtime injects; a new unit test (styles-render-css.test.ts) asserts styles.css carries it verbatim so the two never drift (R0). The runtime (a non-Obsidian bundle) keeps its own <style>.
• Bug 92 — A long-press on an image opened the editing toolbar everywhere, including reading view. A 500 ms touchstart handler called onImageSelected for ANY <img> with no view gate, so on touch a long-press in reading view raised the toolbar (and from there the in-place crop on the .image-embed host) — a double violation of F5/F7 (editing is live-preview only; "no explicit platform-specific trigger"). The click/hover entries were already .markdown-source-view-gated; only this path leaked. Fixed by removing the long-press handler entirely — a plain tap fires a click, which the existing handler already selects in live preview (F7), so touch keeps working. With editing now provably LP-only, the dead reading-view scaffolding it had justified is gone too: the crop host drops its .image-embed branch (always .lie-wrapper), the filter/class panels' pane lookup drops .markdown-reading-view, and the related comments (incl. the native-handle CSS, Bug-32 G) are corrected. No behaviour change in live preview.
• Bug 91 — Decoration / snippet classes + the inline marker rode the <img>, not the outer (the code violated AD2/AD3). render-core's applyClasses put the user .class list + the lie-inline marker on the <img>, which sits inside the wrapper's overflow:hidden — so a decoration class's box-shadow / outset border was clipped (the real defect behind the reverted shadow/border workaround), and the styling needed :has(img.lie-*). Fixed by Decision 28 + Change 36: the user .class list AND the lie-left/right/center / lie-tall / lie-inline markers now ride the OUTER .lie-image-area (box effects no longer clipped). The inline + embed-shrink-wrap :has are gone (direct selectors); the standalone runtime is :has-free; the plugin's alignment-float :has now reads the outer's marker (.host:has(.lie-image-area.lie-…)) — the tolerated CM-context case (Decision 28, the plugin can't class the host it doesn't own), not a leftover defect.
• Bug 90 — Rotating an explicitly-sized image stretched the height instead of swapping the footprint. The toolbar's rotate only flips the rotate field (±90°); the footprint must be derived from rotate + the stored size — which the crop path did (rotatedAabb) but the non-crop path did not. So a width=400 image (height 200 via the intrinsic ratio) kept width=400 on a quarter-turn while the swapped aspect-ratio ballooned the height to 800, instead of becoming width=200 / height=400; same for an explicit width=400 height=200. Pre-existing (the routeBoxStyle / no-explicit-width logic was unchanged by the v0.6.1 review work). Fixed with a pure, unit-tested rotatedFootprint() (renderer-logic) that the non-crop sizing now uses: explicit width+height swap via rotatedAabb; width-only / height-only rotate the base footprint using the intrinsic ratio (the other axis follows the swapped --lie-auto-aspect); an explicit aspect-ratio swaps too. 0°/180° are unchanged; non-px values (e.g. a preset var(--lie-size-…)) pass through.
• Bug 89 — A failed/denied vault write during Export was silently swallowed. The Export fallback modal passed an async callback into a () => void slot, so a rejected adapter.writeBinary (e.g. a permission-denied or invalid path, the Bug 72/R26 write path) was an unobserved rejection — no error surfaced. The callback is now synchronous and threads the write into the outer Promise, so a failure rejects and is caught by exportImage's try/catch → the "Export failed" notice.

[0.6.0] - 2026-06-06

A visual-identity milestone: a refreshed icon set, a new brand mark, hover micro-animations, replace and accessibility features, plus a batch of resolved design decisions. This is a deliberate minor bump (0.5.4 → 0.6.0) — a one-off user decision for this milestone that overrides the project's usual patch-only policy.

• Decision 24 — Crop edge/corner handles should reshape the window from the grabbed side. D8 clarified: the grabbed corner/edge moves while the opposite side stays anchored — corners keep the aspect ratio (aspect-locked), edges are single-axis. The current code stretches the whole image instead, recorded as Bug 80 (deferred — to be batched with future crop-component work).
• Decision 23 — The filter histogram should reflect the LIVE filter (and the panel works on filter=). The panel reads/writes the inner-image filter= attribute (a {style="filter:…"} on the outer box is an expert edge case the editor doesn't author — and the export doesn't bake it, Bug 82). The histogram, though, currently samples the ORIGINAL image — updateHistogram draws the img with no ctx.filter, so the CSS filter (a render effect, not pixel data) never reaches getImageData and the chart never changes as you tune the filter. That's wrong — it should track the live-filtered result → Bug 83.
• Decision 22 — A Reading-view click opens no toolbar — confirmed intended. Reading view is read-only; editing needs the source / Live-Preview editor (the click handler is scoped to .markdown-source-view). Pinned in F7 so it isn't read as a gap.
• Decision 21 — Wikilinks DO carry a caption; link-form parsing must split caption vs. size. A wikilink alias is the markdown-alt equivalent, and |size (now incl. auto and WxH, optionally alongside a delimited caption) shares that slot. On md→wiki the native size folds into the {…} block 1-to-1 (not the pipe) and the caption is preserved; the old code dropped the alias caption — fixed as Bug 81.
• Decision 20 — Shared sub-menu host (AnchoredSubmenu) API documented as shipped. No behaviour change: the placement modes (under-toolbar / beside-image / centered), allowFlip + the pane-bound, the body-builder contract and the ✓/✗/Esc + auto-persist lifecycle are now spelled out in D6/F14.
• Decision 19 — reset() vs. clearStaleTransform both broke the box invariant. Verdict: both were wrong. The 3-layer DOM — wrapper (size/ratio/layout) → inner frame (rotate/flip) → img (filter/crop) — is always present and at least parametrized with native defaults; the box is never empty (or the image vanishes) and the img is never bare (never without a box), only the parametrization varies. So reset() must not leave an empty box (→ Bug 78 facet) and clearStaleTransform must not unwrap to a bare img (→ Bug 79).
• Decision 18 — T2.3 "never width + height together" vs. custom-size emitting both. Resolved by making T2.3 precise rather than changing code. Explicit width + height from the custom-size path is allowed (D6.1/F24); the old "height → aspect-ratio for responsive" idea is moot there because a fixed px width sits alongside the height. T2.3 now reads: presets / auto never co-emit width + height; the explicit custom-size path may.
• Decision 17 — Filter panel always-right vs. D7 "the side with more room". Closed: D7 as written is the wanted behaviour. The panel hard-coded allowFlip: false, contradicting D7 — recorded as Bug 77 rather than reworded.
• Decision 16 — data- prefix on write. Closed: keep bare keys, by design. Pandoc prepends data- itself (HTML5-conformant), the brace-stripper likewise; data- inside {…} would be long, unreadable and glightbox-incompatible. Verified the standalone runtime reads both spellings (bare and data-) via render-core.ts and shared runtime.ts — requirement already met, no bug.
• Decision 15 — F14 listing Export among the live panels. Reworded: Crop / Filter / Resize use the AnchoredSubmenu, but Export is the native OS save dialog (desktop, @electron/remote) / a vault-path modal (mobile, adapter.writeBinary) — Export is not a live panel.
• Decision 14 — Enter = accept globally while a panel is open. Closed by design — confirmed it feels right and fits the modal / active-region model (panel = focus, Enter ↔ ✓, Esc ↔ ✗).
• Decision 13 — Residual display-mode coupling (AD3). Closed by design. Alignment necessarily couples display (center = block + centred, inline = inline) and the explicit px width keeps the footprint identical; documented in AD3 as intended, no refactor.
• Change 32 — Hover micro-animations. Subtle CSS micro-interactions on icon hover/click — rotate-cw winds up, flip tips over, crop snaps, the eraser wiggles, image-down bounces — all gated by prefers-reduced-motion. The reset (eraser) and reset-all (copy-x) share the wiggle so they read as one reset family.
• Change 31 — New brand / plugin icon. A filled mark: a large spark (replacing the sun) over two mountains — left with a plateau (clipped peak, gentle slope), right taller and clipped at the edge, slightly tilted with corner brackets. Wired centrally from docs/img/logo.svg (fixed colour) and src/brand-icon.ts (addIcon, currentColor) into the editing-toolbar submenu, the README, MkDocs (logo + favicon) and the docs index. (No settings-tab header — that would re-introduce the R22 plugin-name heading the project removed in Bug 68. No settings-sidebar nav icon either: community plugins conventionally don't carry one, so adding it would stand out rather than fit in.)
• Change 30 — Icon redesign across the toolbar and editing-toolbar. A refreshed Lucide set: filter sliders-horizontal → blend, custom-size maximize → image-upscale, export download → image-down, CSS-classes chevron-down → braces, inline/block gallery-horizontal-end → wrap-text, the layout trigger layout-list → text-quote, and reset / reset-all to eraser / copy-x. Newer glyphs carry fallbacks (image-upscale → maximize, image-down → download, wrap-text → align-justify, eraser → x-circle).
• Feature 36 — "Button outlines" accessibility setting (Auto / Always / Never). Controls the icon button outlines; Auto keys off prefers-contrast / forced-colors.
• Feature 35 — Replace image / Replace all. Swap the underlying image file for another — a single embed, plus an "all occurrences / whole note" variant; the replace / replace-all icons mirror the reset / reset-all naming. Available as a command and in the editing-toolbar.
• Bug 88 — The CSS-classes picker was a bare absolute-positioned dropdown, not a panel. It was hand-built chrome (a positioned menu / a centered multi-image list), so it shared none of the panel lifecycle. It is now a proper AnchoredSubmenu subpanel (ClassPanel, mirroring FilterPanel): a search box on top + a scrollable class list below, beside-image placement with the toolbar-top anchor (Bug 87) and pane-bound flip (Bug 77), the greyed toolbar, and accept/cancel/Esc + auto-persist. Toggling still writes immediately (the old semantics); multi-image opens the same panel centered and re-resolves locations live per toggle. Search lives in a pure filterClasses (8 new unit tests).
• Bug 87 — Filter panel opened too high / could overlap the window top. In beside-image placement the panel's top was the IMAGE top, then clamped to the panel height — so a tall panel (histogram + presets + sliders) got shoved UP, above the image and even off the window near the top edge. It now anchors its top to the toolbar (a fixed reference just above the image, so the two share a top edge) via a new topAnchorTop arg to placeSubmenu; it slides up ONLY on a window-bottom overflow and snaps back on the next reposition (sticky at toolbar height) — no scrolling, the panel stays whole. Only the filter panel uses beside-image; size/crop (under-toolbar) + the centered multi-image panel are unaffected.
• Bug 85 — Toolbar buttons showed Obsidian's default button frame (a 1px shadow ring + white fill). Obsidian styles a raw <button> via button:not(.clickable-icon) (its background + the --input-shadow inset ring), which out-specified our single-class .lie-toolbar-btn flat rule — so even background: transparent lost and a frame showed. The flat-button rules now repeat the class (.lie-toolbar-btn.lie-toolbar-btn, specificity 0,2,0 > 0,1,1) to kill the native background + shadow, with hover/active bumped to match — the buttons read flat again, as the preview always did.
• Bug 83 — The filter histogram never changed as you tuned the filter. updateHistogram drew the <img> to a sampling canvas with no ctx.filter, so the CSS filter (a render effect, not pixel data) never reached getImageData — the chart always showed the ORIGINAL image (pointless feedback). It now sets the working filter on the sampling context (tempCtx.filter = filterToCss(currentFilter())) before drawImage, so the RGB histogram tracks the sliders live (Decision 23). (The crop half — the histogram should also reflect the cut region — is tracked as Bug 84.)
• Bug 81 — Link-form conversion dropped the wikilink caption and mis-routed the native size. The parser hardcoded caption: "" for wikilinks and read only a numeric tail as size; conversion put the size in the pipe and dropped the caption. Now one shared splitTail parses the alias/alt into a caption and a size — the size grammar follows the obsidian_image_caption convention (WxH with auto, e.g. autox200), and a quoted ("…") caption may sit alongside it. md→wiki folds the native size into the {…} block 1-to-1 (never the pipe) and preserves the caption; captions, Replace and the conversion all share the one parser. (The caption delimiter " is hardcoded for now — a candidate future setting.)
• Bug 79 — clearStaleTransform stripped the 3-layer box. It called unwrapBox, violating the box invariant "the DOM structure is always present, only the parameters change" (Decision 19). It now falls back to native default parameters — the box and the img wrapper stay — instead of unwrapping to a bare img; likely shares a root cause with Bug 78.
• Bug 78 — A cropped image without a width vanished instead of falling back to native sizing. Removing the width of a cropped image collapsed the box. The outer box now follows the same rules as any image — max document width, capped to the original intrinsic size (height or width depending on the box rotation), aspect-ratio derived natively. Crop is not a special case: it only affects the inner <img>, never the box-sizing rules (AD3/R0). The no-width cropped path is now routed through the same native sizing path in buildLayers.
• Bug 77 — The filter panel always docked on the right, not the side with more room (D7). The guided flip was re-enabled: placeSubmenu takes an optional pane bound (default = viewport, so other callers are unchanged), and in beside-image mode it now decides right/left by space within the editor pane (.markdown-source-view / .markdown-reading-view, sidebar excluded), flipping only if the panel fits there — so it never spills over the file explorer (the Bug-56 concern is preserved).
• Bug 66 — Crop handles were hardcoded white (#fff), invisible in light mode. The crop frame, the corner/edge handle squares and the rotate knob + stem are now var(--interactive-accent) (theme- adaptive, so they read in light and dark); the frame is dashed to signal the crop layer (distinct from the solid native resize handle), and the squares carry a --background-primary contrast ring so they stay legible on any image.

[0.5.4] - 2026-06-06

Editing-toolbar integration now actually shows up — as the wanted horizontal icon row.

• Change 29 — Reset command renamed and re-iconed. The single-image reset read "Reset all" with an undo-2 icon — easily confused with the page-scope "Reset all images on this page", and the icon looked like Undo (right next to the real Undo in the editing-toolbar). It's now "Reset image" ("Bild zurücksetzen") with an undo icon, across the hover toolbar, the editing-toolbar submenu and the command palette. (The sub-panel's own "reset this" button keeps undo-2.)
• Bug 75 — Editing-toolbar integration showed nothing. Three causes: the entry used menuType: "dropdown" (a popup, not inline icons); it was appended at the end of a full top bar where it overflowed off-screen; and the live toolbar was never rebuilt after the settings write. It now injects the full image toolbar as a horizontal icon submenu (menuType: "submenu") near the left edge and forces a rebuild, so it appears immediately; on load it self-heals/migrates an existing entry.

[0.5.3] - 2026-06-06

CSS-class management: restore now handles commented-out classes, and the bundled example file gets a full install / reset / uninstall lifecycle.

• Change 28 — Settings reuse Obsidian's own terminology. Feature names (CSS snippets, Appearance) are pulled from Obsidian's own translations (i18next), plus a native info callout explaining the toggles only gate the toolbar picker; the bundled snippet is presented as a built-in CSS snippet, not an "example".
• Feature 30 — Bundled-snippet file lifecycle in settings. The example-snippet entry adapts to the file's state: missing → Install; modified → Reset to shipped (plus Uninstall); unchanged → Uninstall (confirmed first — sterner when the file carries your edits).
• Bug 74 — Per-class restore did nothing for a class "deleted" by commenting it out. The presence check matched the class text inside the /* … */ comment, so it edited the rule in place (still commented) instead of re-adding it; it now judges presence on the comment-stripped CSS.

[0.5.2] - 2026-06-05

Image commands now act on a whole selection of images at once.

• Feature 29 — Standalone panels for multi-image filters / custom size / CSS classes. These open as a centered "N images" panel and apply to every selected image (preview fans out live); crop and export stay single-image.
• Feature 28 — Multi-image commands. When the selection covers two or more images, a command runs on all of them in one undo step (rotate/flip relative per image; align/size set the same value). With no selection it behaves as before on the hover/cursor image; toolbar buttons stay single-image.

[0.5.1] - 2026-06-05

Commands now actually appear in the command palette, and gain a first page-scope command.

• Feature 27 — "Reset all images on this page" command. A page-scope command (always available) that strips every {…} transform block in the active note in one undo step. First of the page-scope command category.
• Bug 73 — Commands were invisible in the command palette. Every command self-gated on a hovered image via checkCallback, which Obsidian hides when false — and opening the palette clears the hover. Image commands now resolve their target from the hovered image or the cursor line.

[0.5.0] - 2026-06-05

Settings-panel rework: the panel is more compact up top and the CSS-classes feature gets a proper management surface.

• Decision 12 — Settings-panel rework forks. "Obsidian off" = no enabled snippet (greys the master toggle + links to Appearance → CSS snippets); the master toggle gates the decoration-class feature only (alignment/inline stay); restore is per-class plus a whole-file "Reset to shipped".
• Change 27 — CSS section reflects Obsidian state. With no CSS snippet enabled at all, the master toggle greys out and an inline notice + link points to Appearance → CSS snippets.
• Change 26 — Native grouped-card settings layout. Related settings sit in Obsidian's own setting-group / setting-items cards (uniform with the core snippets & community-plugins pages).
• Feature 26 — Editing-toolbar integration. Adds the whole image toolbar as one submenu (validated against editing-toolbar 4.0.8); the settings entry adapts to not-installed / installed-but-disabled / enabled.
• Feature 25 — CSS-classes management surface. A master toggle plus a plugin-list-style overview (search, per-file rows, bundled classes pinned), changed/deleted detection vs the shipped snippet, per-class restore, and name-collision warnings.

[0.4.2] - 2026-06-05

Community-directory release-compliance pass — closes the remaining review-checklist rules (R20–R30) so the plugin is ready for submission. No change to image-editing behaviour.

• Change 25 — [Release-Requirement] Community-directory compliance pass. UI-guideline settings tab (no plugin-name heading, setHeading(), CSS class for the warning colour), sentence-case localized command names, normalizePath() on user/constructed paths, the Vault API over the adapter, manifest description, and README disclosures (R20–R30).
• Bug 72 — [Release-Requirement] User/constructed paths not run through normalizePath(). The export fallback's vault path and the constructed snippet paths now go through normalizePath() (R26).
• Bug 71 — [Release-Requirement] Remaining UI headings were Title Case. "CSS snippets" / "Editing toolbar integration" are sentence case (R25).
• Bug 70 — [Release-Requirement] Command names were Title Case and bypassed i18n. Size/align commands read Size: small / Align: left and route through en/de (R24).
• Bug 69 — [Release-Requirement] Raw HTML section headings instead of setHeading() (R23).
• Bug 68 — [Release-Requirement] Plugin-name top-level heading in settings removed (R22).

[0.4.1] - 2026-06-05

• Decision 11 — Docs-site stack: ProperDocs + MaterialX. Single-version site, native version badge; mike and stock MkDocs/mkdocs-material rejected (abandoned / EOL).
• Decision 10 — Snippets install/reset. Install copies shipped CSS without force-overwrite; Reset is per-file to the shipped original; only Obsidian-enabled snippets are scanned/offered.
• Change 24 — The standalone runtime now ships the default decoration classes (rounded / shadow / bordered / circle), extracted to the Obsidian-free src/bundled-snippet.ts shared by plugin and runtime.
• Bug 62 — Click-away now closes an open filter/size panel. The boundary was the whole image+toolbar region (the image fills the canvas, so it stayed open); it now shrinks to the sub-panel, and clicking anywhere else closes + persists. Crop stays exempt.

[0.4.0] - 2026-06-05

A ground-up rework of how images render and how cropping works, a cleaner toolbar / panel interaction, a real test suite and refreshed documentation.

• Decision 9 — Keep the *-logic.ts split. Pure logic stays split from framework-coupled code so it's unit-testable without an obsidian/CM mock.
• Decision 8 — Temperature retired. The virtual-temperature control had no panel row; removed.
• Decision 7 — Routing rule. Image = box; everything routes to the box except transform/filter (→ the image); alignment/inline → the embed.
• Decision 6 — Inline ≠ size; uniform chrome. Inline and size are orthogonal; no "chrome skipped for inline" special case (R0).
• Change 23 — Reveal default = auto (docs aligned: alwaysShowLink off = reveal on hover / active line).
• Change 22 — Pure DRY/KISS refactors — shared locateActiveImage, single nonDefaultFilter, BOX_CLASS, one labelled textButton.
• Change 21 — Dead-code sweep — removed getPreset / setPresetWidth / parseLocationTransform and their tests.
• Change 20 — Portable bare-key format fulfilled — one format, three consumers (no-JS fallback, runtime, writer); runtime-only keys degrade to the original image.
• Change 19 — Obsidian-free render core extracted (render-core.ts: buildLayers + RENDER_CSS, injected by both plugin and runtime — one source, identical render).
• Change 18 — Temperature filter removed — the seven sliders + five presets remain.
• Change 17 — Toolbar + panel are one active region. The bar stays (greyed) as you move onto the panel; the accept (✓) / cancel (✗) icons are back, routed through an explicit exit reason.
• Change 16 — Auto-persist editing. Leaving a panel persists once as a single undo step; the per-panel Reset is the only in-session revert.
• Change 15 — Crop teardown restores all transient overrides on every exit (a single exitCropMode, no second path).
• Change 14 — Crop serialization. Placement = transform=, cut shape = aspect-ratio= (stored only when it differs); never a fixed px height.
• Change 13 — Box rename. The uniform image box is .lie-image-area, the chrome container .lie-box; lie-rotate-box retired.
• Change 12 — Cleaner stored format. Short bare keys ({rotate=90 flip=horizontal width=300}); align / width / filter are real HTML attributes so they survive in other renderers (old blocks still parse).
• Change 11 — Uniform rendering model (AD3) + 3-layer DOM. Every image renders through one outer / frame / image structure, identical in both views.
• Change 10 — Ground-up rework landed. Native CSS storage (transform/filter verbatim), declarative box→image sizing (no JS measure-retry loop), pure-CSS caption, LP overlay + native edit.
• Feature 24 — Portable runtime. A standalone script reproduces the rendering on a published page (MkDocs/Material); align/width/filter already render natively without it.
• Feature 23 — Trackpad rotate gesture (macOS) — rotate in the crop editor with a two-finger twist.
• Feature 22 — In-place crop with edge handles. Edits the image where it sits — 4 corner (aspect-locked) + 4 edge handles + a rotate knob; what you see is what gets committed.
• Bug 64 — Group popups / class dropdown now couple to the active region (kept visible, not greyed).
• Bug 63 — Panel visibility firmly coupled to toolbar visibility — no un-greyed flash while a panel is open.
• Bug 61 — Toolbar ↔ sub-modal are one active region — the bar no longer drops the instant the pointer leaves the image rect.
• Bug 60 — Auto-persist on anchor-disconnect wrote the right occurrence of a duplicated image (uses the location captured at panel-open, not a basename scan).
• Bug 59 — The "icon" size preset now renders the image inline as an icon, not merely at icon height.
• Bug 58 — A filter-only image is now hydrated by the portable runtime ([filter] added to the claim selector).
• Bug 57 — Reading-view render of a duplicated image now resolves the n-th occurrence, not the first.
• Bug 56 — Toolbar/menu edits wrote to the WRONG image. The resolver matched a basename's first occurrence; it now resolves the exact line by DOM position (posAtDOM).
• Bug 55 — The {…} block regained its syntax highlighting (marked as a cm-url string).
• Bug 54 — <> dismiss hides the whole raw embed (the fake link and the {…}).
• Bug 53 — The reveal toggle shows the <> glyph again, not an eye icon.
• Bug 52 — Crop pan grabs the whole image (inside and outside the cut frame).
• Bug 51 — Crop editor migrated to the live 3-layer model — true in-place editing, centre origin, edge handles, width-resize preserves the crop; preview == committed.
• Bug 50 — rotate/flip no longer drifts an already-cropped image (orientation rides the frame about its centre; the crop placement is untouched).
• Bug 48 — Reading-view sizing of transformed images — a transformed box caps to the column and a floated image's wrap text starts at the image top (matches live preview).
• Bug 45 — Latent box-selector bug — previewSize queried a non-existent class on rotated images.
• Bug 44 — The demo notes migrated to native syntax.
• Bug 43 — Resize handle CSS.
• Bug 42 — Size "Original" / a cleared field no longer collapses the box.
• Bug 41 — The filter panel gained the shared per-panel reset.
• Bug 40 — The crop overlay is exempt from the dismiss handler.
• Bug 39 — The crop committed result equals the framed region.
• Bug 38 — Crop rebuilt: handles scale the inner image toward the frame centre.
• Bug 36 — Toolbar anchored to the image top via the box.
• Bug 35 — Captions below the image — centred, width-limited, Markdown-rendered, pure CSS.
• Bug 34 — Snippet "png" came from a CSS comment — strip comments + filter file extensions.
• Bug 33 — Rotate centred via a translate(-50%,-50%) prepend.
• Bug 32 — Reset no longer whites-out the window — an empty class token in the CM update cycle is guarded.

[0.3.0] - 2026-06-04

A live-preview rework: images can float and wrap text, mid-paragraph images become editable, and the raw-link reveal is simpler and calmer.

• Decision 5 — RevealMode = auto|always is not a retired mode cycle. The two values derive from the global default-state setting; the <> dismiss is a separate per-line override.
• Decision 4 — The floating toolbar and the in-image toolbar are not a duplicate. Both build via buildToolbarElement; only host + positioning differ.
• Decision 3 — Reveal/edit model: overlay + CSS reveal + native edit. The LP adapter never replaces the line; Obsidian renders its CSS-suppressed native embed and gives the cursor-reveal, the plugin overlays its own image.
• Change 9 — Standalone vs. inline placement. An image on its own line is a block; one carrying a {…} block stays inline so it can float.
• Change 8 — Column-capped images. A transformed image never overflows its column; it caps to the text width and stays responsive.
• Change 7 — Reveal is now pure-CSS. The raw {…} link shows/hides via CSS keyed on hover and the cursor line, so it no longer fights the editor or churns the undo history.
• Change 6 — Link reveal collapsed to two natural modes (auto / always) from one setting, plus a transient <> dismiss that auto-clears — no tri-state cycle to remember.
• Change 5 — 2026-06-02 DRY/KISS audit points dissolved by the rework (no canvasFilter table, shared rotatedAabb, a single buildSlider, pure-CSS caption).
• Change 4 — CLAUDE.md slimmed to a lean build/debug guide; the duplicated requirements/architecture/known-bugs were removed.
• Feature 21 — Native-look resize handle — the resize corner matches Obsidian's native image handle.
• Feature 20 — Float & text-wrap in live preview — left/right-aligned images float and wrap text in LP; a tall float (>~250 px) stacks as a block.
• Feature 19 — Inline image embeds — a mid-paragraph image gets the same toolbar/transforms and flows inline.
• Bug 49 — Inline-icon / tiny-image toolbar mis-positioned — the floating bar now sits above the image and floats out by coverage.
• Bug 47 — Tall float (>~250 px) derenders on scroll in LP — capped (stacks as a block in safe mode, both views).
• Bug 46 — Live-preview float breaks CM6 layout — resolved via an inline non-BFC float-escape widget (multi-line wrap, zero height desync, image stays clickable).
• Bug 37 — Toolbar fold-then-wrap — a measured reflow folds groups to a submenu, then wraps at the dividers.
• Bug 31 — No native <> edit-block icon leak — the native edit-block button is suppressed.
• Bug 30 — Reveal/edit sit above the image (the AD5 overlay).
• Bug 29 — Reveal toggles work (the true AD5 overlay model, after a brief mis-build was reverted).
• Bug 28 — Scroll jank; image sections rendered very late — the block widget gained an estimatedHeight derived from a pure estimate.
• Bug 27 — lie-center only centred on hover — centre via text-align:center on a full-width block embed (no !important arms race).
• Bug 26 — Inline (mid-text) image rendered native & full-size — now the same widget in an inline mode.
• Bug 25 — A resized crop left an empty band (the caption was pushed below) — the box is aspect-correct when one dimension is given.
• Bug 24 — Standalone classes lost in live preview — the {…} braces were passed to the parser, dropping the leading .class; strip the braces.

[0.2.0] - 2026-06-01

The second release: new captions and a proper export save dialog, a reworked toolbar layout, and a round of fixes from in-Obsidian verification of the initial version.

• Change 3 — Image spacing. Stacked images use Obsidian's native 6px vertical spacing.
• Change 2 — AUTO link reveal also appears when the editor cursor is on the image's line, in addition to on hover.
• Change 1 — Export fidelity. Export renders every transform exactly as displayed — size, rotation, flip, crop and filters.
• Feature 18 — Per-panel reset — the size and crop sub-menus each get their own Reset, separate from the toolbar's reset-all.
• Feature 17 — Export save dialog — saving an export opens the OS-native save dialog at the original's folder, name pre-filled {original}-{n}; nothing is ever overwritten silently.
• Feature 16 — Image captions — an optional setting renders the image's alt text as a caption below the image (Markdown supported), off by default.
• Bug 23 — A wikilink caption ending in a size token (![[img|caption|300]]) now reads the same in both views.
• Bug 22 — Exporting a filename with a literal % no longer fails; the name-probe loop is bounded.
• Bug 21 — The filter panel closes itself when its anchor leaves the DOM (no orphaned panel / leaked listeners).
• Bug 20 — Reading-view captions no longer re-render on every editor update when the text is unchanged.
• Bug 19 — The crop sub-menu's Reset restores the full image when re-cropping an already-cropped image.
• Bug 18 — Removing an alignment/decoration class takes effect immediately in reading view (tracked via data-lie-classes, cleared on reset).
• Bug 17 — A height-only resize now keeps its aspect ratio when exported (no stretch).
• Bug 16 — The custom-size live preview works on plain (un-rotated/cropped) images again (the display:contents box had swallowed the size).
• Bug 15 — The resize / selection frame now hugs the image instead of sitting offset.
• Bug 14 — An image with no explicit size no longer overflows its column.
• Bug 13 — The revealed link editor is borderless again.
• Bug 12 — Export no longer fails with "file already exists" (superseded by the new save dialog).
• Bug 11 — Resize affordance restored — uses Obsidian's native handle, shown on toolbar hover, hidden in crop.
• Bug 10 — Alignment classes (left / center / right) now actually float / align the image.
• Bug 9 — The custom-size sub-menu gained a height field (width + height side by side).
• Bug 8 — The temperature slider's own thumb stays put (sliders matched by data-key).
• Bug 7 — Filter-panel slider rows no longer overlap.
• Bug 6 — Toolbar icons grouped with dividers (and divider-wrapping).
• Bug 5 — The + / − size buttons were dropped (resize via the handle + custom-size).
• Bug 4 — The crop editor no longer breaks when the image is dragged inside it.
• Bug 3 — The filter panel docks beside the image, follows on scroll, hides offscreen.
• Bug 2 — Quarter-turn-rotated images size their reflow box correctly (a single render path).
• Bug 1 — AUTO link reveal now shows from the first render.

[0.1.0] - 2026-06-01

First public release. Non-destructive image editing for Obsidian: a hover toolbar lets you crop, rotate, flip, resize and filter images, all live — the original file is never modified. Edits are stored in a portable trailing {…} attribute block, so the same notes still render correctly when published via MkDocs/Material.

• Decision 2 — Crop pixel-quantization. The cut quantizes to whole pixels + fixed angle steps live during the interaction.
• Decision 1 — Export resolution. Export renders from the original image's native resolution; display size never reduces it.
• Feature 15 — Settings tab — toggle the hover toolbar, manage detected snippet classes, optional (off-by-default, version-gated) editing-toolbar integration.
• Feature 14 — Commands — context-aware commands (active only with an image in context) for rotate, flip, crop, filters, sizing, alignment, add-class, reset, custom-size, toggle-inline, export.
• Feature 13 — Follows Obsidian — markdown vs. wikilink form follows the central "Use [[Wikilinks]]" setting; the UI follows Obsidian's locale (English fallback, German included).
• Feature 12 — <> link reveal — shows the raw link as editable text above the image and writes edits back live.
• Feature 11 — Anchored sub-menus — size, crop and filter controls share one component: the toolbar greys out, confirm / cancel are icons, Esc cancels.
• Feature 10 — Hover toolbar — native Lucide icons, grouped Edit/Layout buttons that collapse into an overflow sub-menu when space is tight, plus mobile long-press.
• Feature 9 — Both views — rotation, flip, filters, crop and sizing render identically in reading view and live preview.
• Feature 8 — Vault-snippet classes — image CSS classes discovered from .obsidian/snippets/ and offered in a dropdown; each individually de-selectable, refreshed on file changes.
• Feature 7 — Preset classes — built-in toggleable classes for size, alignment and decoration, with a one-click reset.
• Feature 6 — Export — render all transforms and filters to a new {original}-edited.{ext} file via canvas; the original stays untouched.
• Feature 5 — Filters — brightness, contrast, saturation, hue-rotate, blur, grayscale and sepia sliders, a live RGB histogram, and the presets Sepia / B&W / Vintage / Cool / Warm.
• Feature 4 — Resize — drag the native corner handle (aspect kept) or open the custom-size sub-menu for an exact width / height and the quick sizes small / medium / large / original.
• Feature 3 — Crop — an in-place crop editor with a fixed, resizable frame over a movable, rotatable and scalable original; free rotation snaps to whole pixels + 0.1° steps, plus aspect presets 16:9, 4:3, 1:1.
• Feature 2 — Rotate & flip — quarter-turn rotation (CW / CCW) and horizontal / vertical flip; rotated images keep the same content width and reflow correctly.
• Feature 1 — Non-destructive editing — all transforms are written to a trailing {…} block; the image file, alt text, path and native |size are never touched.