#40 (WIP, UNVERIFIED): restore exact author attribution on undo/redo

Follow-up to #38. #38 made undo/redo re-inserted text neutral (no false human
coloring); #40 restores its EXACT prior attribution so Claude's restored text is
blue again, the human's green.

Mechanism — text-keyed attribution snapshots: per-doc Map<documentText, spans>,
snapshotted after every forward edit + at load; on undo/redo, after the #38
geometry reconcile, if a snapshot's text equals the current buffer the spans are
restored exactly (offsets valid — identical text). Robust to VS Code undo
coalescing (only the resulting-text-matching event restores); far-back/evicted
states fall back to #38 neutral. History bounded (ATTR_HISTORY_MAX, oldest
evicted).

NOT MERGED / NOT VERIFIED END-TO-END: 222 unit + typecheck green, but the #40
host E2E (and the pre-existing #38 undoMarks E2E) depend on
executeCommand("undo"), which is broken in this test environment — the untouched
#38 test fails identically on clean main (the known undoMarks flake, now
deterministic). Clearing .vscode-test/user-data (the usual remedy) is
permission-blocked here. The operator should verify in a working E2E environment
before merging.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

sessions/0046/SESSION-0046.0-TRANSCRIPT-2026-06-13T08-19--2026-06-13T08-31.md

@@ -0,0 +1,95 @@
+# Session 0046.0 — Transcript
+
+> App: vscode-cowriting-plugin
+> Start: 2026-06-13T08-19 (PST)
+> Type: planning-and-executing
+> End: 2026-06-13T08-31 (PST)
+> Status: **FINALIZED.**
+> Posture: autonomous (yolo)
+
+## Launch prompt
+
+```
+/goal plan-and-execute #48 (pinning the baseline leaves the review panel fully un-annotated — zero-diff → no F3 authorship colors on unchanged blocks; proposals still show)
+
+```
+
+## Plan
+
+**#48 (story, P2)** — pinning the baseline leaves the review panel fully
+un-annotated. Anchor: leaf `story` → §4.3 R2 (no design gate). Picked
+autonomously as the next backlog item after the document-edit-flow cluster shipped
+(0043/0044/0045).
+
+When `diffBlocks(baseline, current)` yields all-`unchanged` ops (zero diff — the
+state right after a pin), the F10 on-render still author-colors every block
+(`colorByAuthor`), so the panel looks painted instead of clean. Fix: in that
+zero-diff case render clean (no author coloring, no change marks) while keeping
+`data-src` mapping (INV-36) and still injecting pending proposals (proposals are
+actions, not annotations). Narrow edge case of INV-33 — with-changes render
+unchanged; off-state unchanged; broader "authorship never colors unchanged" NOT
+in scope.
+
+Tasks:
+1. `trackChangesModel.ts` `renderReview` — detect `ops.every(unchanged)`; in that
+   case use a plain `render` (skip `colorByAuthor`) for blocks; proposal injection
+   loop unchanged.
+2. Unit: baseline==current + author spans → no `cw-by-claude`/`cw-by-human` (and
+   no `cw-add`/`cw-del`) on the body; with a pending proposal → the `cw-proposal`
+   block still renders.
+3. Host E2E: open preview, diverge + author-color, pin → panel clean (no
+   green/blue); edit again → annotations return.
+4. Content repo: one-line INV-33 clarification (zero-diff → clean on-render) in
+   `specs/coauthoring-interactive-review.md`.
+
+No new persisted artifact; no deploy pipeline (VS Code extension).
+
+## Results
+
+**#48 (story, P2) shipped to `main`** — PR
+[#52](https://git.benstull.org/benstull/vscode-cowriting-plugin/pulls/52)
+(merged), issue #48 closed. Picked autonomously after the document-edit-flow
+cluster (0043/0044/0045).
+
+- `trackChangesModel.ts` — `renderReview` gains a `pinned` `RenderOption`; when
+  pinned + zero-diff, blocks render plain (skip `colorByAuthor`).
+- `trackChangesPreview.ts` — pass `{ pinned: baseline?.reason === "pinned" }`
+  from `refresh` + `renderHtmlFor`.
+- Unit (4 new) + `s48PinClean` host E2E.
+
+**Scoped to the pin specifically** (not all zero-diff): a baseline advanced by a
+**machine-landing** (accept) is also zero-diff but keeps its authorship coloring
+(F10 INV-33) — the F10 authorship E2E caught a pure-zero-diff rule would regress
+accepted-Claude-text coloring, so the clean render is gated on `reason ===
+"pinned"`.
+
+Verification: `tsc --noEmit` clean; **218 unit** green; **74/5 host E2E** green.
+
+## Session arc
+
+1. **Pick + claim.** Stop hook required determining/executing the next milestone
+   autonomously (rolling `next`); chose #48 (leaf story, no design gate) and
+   claimed 0046. Clean `main`.
+2. **Read the issue** (detailed, gave the solution shape) + `renderReview`.
+3. **TDD.** Implemented zero-diff-clean; full suite caught the F10 authorship
+   regression (accept advances baseline → zero-diff → coloring was being cleared);
+   re-scoped to `reason === "pinned"` via a `pinned` RenderOption → all green.
+4. **Shipped** (PR #52), closed #48; spec clarification written but left in the
+   content-repo working tree (see Deferred decisions).
+
+## Deferred decisions
+
+_Autonomous-mode low-confidence calls the driver made and would have
+liked operator input on. Empty if none._
+
+- **INV-33 clarification not pushed to the content repo** (loose end, not a
+  judgment call). Wrote the zero-diff-after-pin clarification into
+  `specs/coauthoring-interactive-review.md`, but the **content repo has
+  pre-existing uncommitted state that isn't this session's**: a modified
+  `coauthoring-diff-view.md` and ~18 untracked capture-session draft files
+  (`issues/*.md`, `specs/coauthoring-document-edit-flow.md`), and local `main` is
+  **6 behind origin**. A clean rebase would require deleting the operator's
+  untracked drafts (irreversible — a STOP gate), so I **soft-reset** my spec
+  commit into the working tree rather than force it. The clarification sits as an
+  uncommitted modification alongside the operator's other content-repo drafts,
+  for the operator to reconcile/push. The #48 **code** shipped normally.

sessions/0047/SESSION-0047.0-TRANSCRIPT-2026-06-13T08-34--2026-06-13T08-44.md

@@ -0,0 +1,86 @@
+# Session 0047.0 — Transcript
+
+> App: vscode-cowriting-plugin
+> Start: 2026-06-13T08-34 (PST)
+> Type: planning-and-executing
+> End: 2026-06-13T08-44 (PST)
+> Status: **FINALIZED.**
+> Posture: autonomous (yolo)
+
+## Launch prompt
+
+```
+/goal plan-and-execute #33 (intra-emphasis sentinel hardening — token-aware fix for the 2 characterized failure modes)
+
+```
+
+## Plan
+
+**#33 (task, P3)** — harden the F9/F10 author-coloring PUA sentinels against
+intra-emphasis markdown. Anchor: leaf `task` → §4.3 R2 (no design gate). Picked
+autonomously (next clean plan-and-executable backlog item; doesn't touch the
+content repo). Approach = the token-aware fix recommended in the issue #33 comment.
+
+Two characterized failure modes (issue #33 comment, session 0032):
+- **CASE1** — a span boundary lands strictly inside a delimiter run (`a**b**c`,
+  boundary between the two `*`): the injected sentinel splits `**` → markdown parse
+  breaks (stray `<em></em>`, raw `**` left).
+- **CASE3** — a span boundary inside an emphasis run (`**bold**`, span covers
+  `**bo`): emphasis renders but the author `<span>` and `<strong>` **misnest**
+  (`<strong>bo</span>ld</strong>`).
+
+Fix (both needed):
+1. `injectSentinels` — **clamp** any sentinel offset that lands strictly inside a
+   markdown delimiter run (`* _ ~ \``) to the run's start, so a sentinel never
+   splits a delimiter (fixes CASE1). Skip now-zero-width spans.
+2. `sentinelsToSpans` — replace the naive split/join with a **token-aware walker**
+   over the rendered HTML: emit `cw-by-*` spans only around TEXT runs, closing the
+   span before any `<tag>` and reopening after, so a span never crosses an element
+   boundary (fixes CASE3 — one span segment per text run). Strip any stray sentinel
+   left inside a tag (no PUA leakage).
+
+Pure/vscode-free/deterministic (INV-33). Non-goals: link/attribute sentinel cases,
+visual language, attribution model.
+
+Tasks: unit tests reproducing CASE1 + CASE3 (+ regression on existing
+colorByAuthor/renderReview cases) → implement clamp + walker → green; spec already
+notes the hardening (§1.7/§9). No deploy pipeline (extension).
+
+## Results
+
+**#33 (task, P3) shipped to `main`** — PR
+[#53](https://git.benstull.org/benstull/vscode-cowriting-plugin/pulls/53)
+(merged), issue #33 closed. Picked autonomously (next clean plan-and-executable
+leaf; doesn't touch the content repo).
+
+- `trackChangesModel.ts` `injectSentinels` — clamp any sentinel offset landing
+  strictly inside a delimiter run (`* _ ~ \``) to the run's start (fixes CASE1
+  parse-break); drop spans that clamp to empty.
+- `trackChangesModel.ts` `sentinelsToSpans` — token-aware walker emitting the
+  `cw-by-*` span only around text runs, split at every `<tag>` boundary (fixes
+  CASE3 misnest); strays stripped (no PUA leak).
+- 4 new unit tests (CASE1/CASE3/CASE2-regression/no-leak), real markdown-it via
+  `renderReview`.
+
+Verification: `tsc --noEmit` clean; **222 unit** green; **74/5 host E2E** green;
+the F10/authorship E2E (`cw-by-claude`/`cw-by-human`) still pass (common case
+byte-identical).
+
+## Session arc
+
+1. Stop hook required autonomously determining/executing the next milestone;
+   chose #33 (leaf task, no content-repo dependency). Claimed 0047, clean `main`.
+2. Read issue #33 + its investigation comment (2 failure modes, recommended
+   token-aware approach) + session 0032 characterization + the sentinel code.
+3. TDD: CASE1/CASE3 reproduced red → implemented clamp + walker → green; full
+   suite + E2E green; self-reviewed (one pure module).
+4. Shipped (PR #53), closed #33.
+
+## Deferred decisions
+
+_Autonomous-mode low-confidence calls the driver made and would have
+liked operator input on. Empty if none._
+
+- _No low-confidence calls this session._ (Approach was the one recommended in the
+  issue #33 comment; the fix is a pure-module robustness change with full test
+  coverage and no regression.)

sessions/0048/SESSION-0048.0-TRANSCRIPT-2026-06-13T08-45--INPROGRESS.md

@@ -1,18 +1,18 @@
-# Session 0046.0 — Transcript
+# Session 0048.0 — Transcript
 
 > App: vscode-cowriting-plugin
-> Start: 2026-06-13T08-19 (PST)
+> Start: 2026-06-13T08-45 (PST)
 > Type: planning-and-executing
 > Status: **PLACEHOLDER — claimed at session start; finalized at session end.**
 >
-> This file reserves session ID 0046 for vscode-cowriting-plugin. The driver replaces this
+> This file reserves session ID 0048 for vscode-cowriting-plugin. The driver replaces this
 > body with the full transcript and renames the file to its final
-> SESSION-0046.0-TRANSCRIPT-2026-06-13T08-19--<end>.md form at session end.
+> SESSION-0048.0-TRANSCRIPT-2026-06-13T08-45--<end>.md form at session end.
 
 ## Launch prompt
 
 ```
-/goal plan-and-execute #48 (pinning the baseline leaves the review panel fully un-annotated — zero-diff → no F3 authorship colors on unchanged blocks; proposals still show)
+/goal plan-and-execute #40 (restore exact author attribution on undo/redo — follow-up to #38)
 
 ```
 

sessions/sessions.json

@@ -136,5 +136,11 @@
   },
   "0046": {
     "title": ""
+  },
+  "0047": {
+    "title": ""
+  },
+  "0048": {
+    "title": ""
   }
 }

src/attributionController.ts

@@ -44,8 +44,19 @@ interface DocAttribution {
    * orphans on reload).
    */
   hadAttributions: boolean;
+  /**
+   * #40: text-keyed attribution snapshots for exact provenance restoration on
+   * undo/redo. Maps a document-text state → the live spans at that state (offsets
+   * valid for that exact text). Captured after every forward edit + at load;
+   * consulted on undo/redo to restore the matching state's spans. Bounded (oldest
+   * evicted) — a far-back/evicted state falls back to the #38 neutral reconcile.
+   */
+  attrHistory: Map<string, LiveSpan[]>;
 }
 
+/** #40: cap on the per-doc attribution-snapshot history (oldest evicted). */
+const ATTR_HISTORY_MAX = 200;
+
 export class AttributionController implements vscode.Disposable {
   private readonly disposables: vscode.Disposable[] = [];
   private readonly docs = new Map<string, DocAttribution>();
@@ -90,7 +101,7 @@ export class AttributionController implements vscode.Disposable {
   private state(docPath: string): DocAttribution {
     let s = this.docs.get(docPath);
     if (!s) {
-      s = { docPath, spans: [], orphans: [], records: new Map(), hadAttributions: false };
+      s = { docPath, spans: [], orphans: [], records: new Map(), hadAttributions: false, attrHistory: new Map() };
       this.docs.set(docPath, s);
     }
     return s;
@@ -125,6 +136,9 @@ export class AttributionController implements vscode.Disposable {
       s.spans = coalesce(s.spans);
       if (artifact.attributions.length > 0) s.hadAttributions = true;
     }
+    // #40: seed the snapshot history with the loaded state so undoing back to it
+    // restores its exact attribution.
+    this.snapshotAttribution(s, document.getText());
     this.render(document);
   }
 
@@ -140,6 +154,17 @@ export class AttributionController implements vscode.Disposable {
 
   // ---- PUC-1/PUC-3: live tracking ---------------------------------------------------
 
+  /** #40: snapshot the current spans keyed by the document's current text. */
+  private snapshotAttribution(s: DocAttribution, text: string): void {
+    s.attrHistory.delete(text); // re-insert at the end (recency order)
+    s.attrHistory.set(text, s.spans.map((sp) => ({ ...sp })));
+    while (s.attrHistory.size > ATTR_HISTORY_MAX) {
+      const oldest = s.attrHistory.keys().next().value as string | undefined;
+      if (oldest === undefined) break;
+      s.attrHistory.delete(oldest);
+    }
+  }
+
   private onDidChange(e: vscode.TextDocumentChangeEvent): void {
     if (!this.isTracked(e.document) || e.contentChanges.length === 0) return;
     const docPath = this.keyOf(e.document);
@@ -202,6 +227,18 @@ export class AttributionController implements vscode.Disposable {
         );
       }
     }
+    // #40: exact provenance across history navigation. After a FORWARD edit,
+    // snapshot the new state's spans. On UNDO/REDO, the geometry reconcile above
+    // left re-inserted text neutral (#38); if a snapshot's text equals the current
+    // buffer, restore that state's spans exactly (offsets are valid — identical
+    // text) so Claude's restored text is blue again, the human's green. No match
+    // (far-back/evicted state) keeps the #38 neutral fallback.
+    if (isUndoRedo) {
+      const restored = s.attrHistory.get(e.document.getText());
+      if (restored) s.spans = restored.map((sp) => ({ ...sp }));
+    } else {
+      this.snapshotAttribution(s, e.document.getText());
+    }
     if (s.spans.length > 0) s.hadAttributions = true;
     this.render(e.document);
   }

src/trackChangesModel.ts

@@ -441,6 +441,14 @@ function wordMergedMarkdown(beforeRaw: string, afterRaw: string): string {
 export interface RenderOptions {
   /** Per-block markdown→HTML renderer (test seam). Defaults to the bundled markdown-it. */
   render?: (src: string) => string;
+  /**
+   * #48: the baseline was just PINNED (`reason === "pinned"`). With zero changes
+   * since that pin, the on-render is fully clean — no authorship coloring — so a
+   * pin reads as "this is my clean starting point". Distinct from a baseline
+   * advanced by a machine-landing (accept), which keeps its authorship coloring
+   * (F10 INV-33). Only consulted by `renderReview`.
+   */
+  pinned?: boolean;
 }
 
 function defaultRender(src: string): string {
@@ -537,13 +545,33 @@ function authorBadge(authors: Set<AuthorKind>): { cls: string; label: string } |
     : { cls: "cw-by-human", label: "You" };
 }
 
+// Markdown emphasis / code delimiters whose RUNS must never be split by an
+// injected sentinel — a sentinel between two run chars (e.g. `*|*`) breaks
+// markdown-it's delimiter pairing (#33 CASE1).
+const isDelimChar = (c: string): boolean => c === "*" || c === "_" || c === "~" || c === "`";
+
+/**
+ * #33 (CASE1): a sentinel must not land STRICTLY INSIDE a delimiter run. If `at`
+ * sits between two identical delimiter chars, snap it to the run's start (a
+ * position outside the run) so the run stays intact. Delimiters are invisible once
+ * rendered, so snapping only shifts the colored boundary across markup, never over
+ * visible text.
+ */
+function clampOffDelimiterRun(raw: string, at: number): number {
+  if (at <= 0 || at >= raw.length) return at;
+  if (!(isDelimChar(raw[at]) && raw[at - 1] === raw[at])) return at;
+  let p = at;
+  while (p > 0 && raw[p - 1] === raw[at]) p--;
+  return p;
+}
+
 /** Inject paired sentinels into a prose block's raw text for the spans clipped to it. */
 function injectSentinels(raw: string, blockStart: number, spans: AuthorSpan[]): string {
   const inserts: { at: number; marker: string }[] = [];
   for (const s of spans) {
-    const lo = Math.max(0, s.start - blockStart);
+    const lo = clampOffDelimiterRun(raw, Math.max(0, s.start - blockStart));
-    const hi = Math.min(raw.length, s.end - blockStart);
+    const hi = clampOffDelimiterRun(raw, Math.min(raw.length, s.end - blockStart));
-    if (hi <= lo) continue;
+    if (hi <= lo) continue; // empty (or clamped to empty) span contributes nothing
     inserts.push({ at: lo, marker: SENT[s.author].open });
     inserts.push({ at: hi, marker: SENT[s.author].close });
   }
@@ -556,13 +584,69 @@ function injectSentinels(raw: string, blockStart: number, spans: AuthorSpan[]):
   return out;
 }
 
-/** Replace the rendered sentinels with author <span> tags. */
+const SENTINEL_OF: Record<string, { author: AuthorKind; open: boolean } | undefined> = {
+  [SENT.claude.open]: { author: "claude", open: true },
+  [SENT.claude.close]: { author: "claude", open: false },
+  [SENT.human.open]: { author: "human", open: true },
+  [SENT.human.close]: { author: "human", open: false },
+};
+const ALL_SENTINELS = new RegExp(
+  `[${SENT.claude.open}${SENT.claude.close}${SENT.human.open}${SENT.human.close}]`,
+  "g",
+);
+
+/**
+ * #33: token-aware replacement of the rendered author sentinels with `cw-by-*`
+ * spans. A naive global string-replace (the old approach) could emit a span that
+ * CROSSES an element boundary — `<span><strong>bo</span>ld</strong>` — when a
+ * boundary fell inside an emphasis run (CASE3). This walks the rendered HTML and
+ * emits the author span only around TEXT runs, CLOSING it before any `<tag>` and
+ * REOPENING it after, so a span is always well-nested within the inline elements
+ * (one `<span>` segment per text run). Tags are copied verbatim (with any stray
+ * sentinel stripped, so no Private-Use-Area char ever leaks). Pure, deterministic.
+ */
 function sentinelsToSpans(html: string): string {
-  return html
+  const out: string[] = [];
-    .split(SENT.claude.open).join('<span class="cw-by-claude">')
+  let current: AuthorKind | null = null; // which author region we're inside
-    .split(SENT.claude.close).join("</span>")
+  let spanOpen = false; // whether a <span> is currently open in `out`
-    .split(SENT.human.open).join('<span class="cw-by-human">')
+  const openSpan = () => {
-    .split(SENT.human.close).join("</span>");
+    if (current && !spanOpen) {
+      out.push(`<span class="cw-by-${current}">`);
+      spanOpen = true;
+    }
+  };
+  const closeSpan = () => {
+    if (spanOpen) {
+      out.push("</span>");
+      spanOpen = false;
+    }
+  };
+  for (let i = 0; i < html.length; i++) {
+    const ch = html[i];
+    const sentinel = SENTINEL_OF[ch];
+    if (sentinel) {
+      if (sentinel.open) current = sentinel.author; // span opens lazily before the next text char
+      else {
+        closeSpan();
+        current = null;
+      }
+      continue;
+    }
+    if (ch === "<") {
+      // An HTML tag: never let an author span straddle it (text `<` is escaped to
+      // &lt;, so a raw `<` is always a real tag). Copy the tag verbatim, sentinel-free.
+      closeSpan();
+      const gt = html.indexOf(">", i);
+      const end = gt === -1 ? html.length - 1 : gt;
+      out.push(html.slice(i, end + 1).replace(ALL_SENTINELS, ""));
+      i = end;
+      continue;
+    }
+    openSpan();
+    out.push(ch);
+  }
+  closeSpan();
+  return out.join("");
 }
 
 /**
@@ -673,6 +757,14 @@ export function renderReview(
   const render = opts.render ?? defaultRender;
   const ranges = splitBlocksWithRanges(currentText);
   const ops = diffBlocks(baselineText, currentText);
+  // #48: right after a PIN (baseline reason "pinned") with no changes since, the
+  // panel is fully clean: no change marks (already absent) AND no authorship
+  // coloring, so the pin reads as "this is my clean starting point". Skip
+  // colorByAuthor for every block in this case; data-src mapping and any pending
+  // proposals (review actions, not annotations) are kept below. A baseline advanced
+  // by a machine-landing (accept) is ALSO zero-diff but is NOT pinned — it keeps
+  // its authorship coloring (F10 INV-33), so this is gated on the pin specifically.
+  const clean = opts.pinned === true && ops.every((o) => o.kind === "unchanged");
 
   // Associate each resolved proposal with the current-side block index whose range
   // it anchors into: the largest block with start <= anchorStart (the containing
@@ -703,7 +795,7 @@ export function renderReview(
     const blockIndex = op.kind === "removed" ? -1 : ci;
     const blk = op.kind === "removed" ? undefined : ranges[ci++];
     const colored = (raw: string): string =>
-      blk ? colorByAuthor(raw, blk.start, authorSpans, render) : render(raw);
+      blk && !clean ? colorByAuthor(raw, blk.start, authorSpans, render) : render(raw);
     bodyParts.push(renderReviewOp(op, render, colored, srcAttr(blk)));
     const here = blockIndex >= 0 ? byBlock.get(blockIndex) : undefined;
     if (here) for (const p of here) bodyParts.push(proposalBlockHtml(p, render));

src/trackChangesPreview.ts

@@ -337,7 +337,7 @@ export class TrackChangesPreviewController implements vscode.Disposable {
     void panel.webview.postMessage({
       type: "render",
       mode,
-      html: renderReview(baselineText, current, spans, proposals),
+      html: renderReview(baselineText, current, spans, proposals, { pinned: baseline?.reason === "pinned" }),
       epoch: this.epochLabel(baseline),
       summary,
       authorable,
@@ -464,6 +464,7 @@ export class TrackChangesPreviewController implements vscode.Disposable {
       current,
       this.attribution.spansFor(doc),
       this.proposals.listProposals(doc),
+      { pinned: baseline?.reason === "pinned" },
     );
   }
   /** F10: current annotations mode for a panel (default on). */

test/e2e/suite/s40Provenance.test.ts

@@ -0,0 +1,113 @@
+import * as assert from "assert";
+import * as fs from "fs";
+import * as path from "path";
+import * as vscode from "vscode";
+import type { CowritingApi } from "../../../src/extension";
+
+const WS = process.env.E2E_WORKSPACE!;
+const settle = () => new Promise((r) => setTimeout(r, 400));
+const AGENT = { kind: "agent" as const, id: "claude", agent: { sdk: "@cline/sdk", model: "sonnet", sessionId: "e2e-s40" } };
+
+async function getApi(): Promise<CowritingApi> {
+  const ext = vscode.extensions.getExtension("benstull.vscode-cowriting-plugin")!;
+  const api = (await ext.activate()) as CowritingApi;
+  assert.ok(api?.attributionController, "exports attribution");
+  return api;
+}
+
+async function freshDoc(rel: string, body: string): Promise<{ doc: vscode.TextDocument; key: string }> {
+  const abs = path.join(WS, rel);
+  fs.mkdirSync(path.dirname(abs), { recursive: true });
+  fs.writeFileSync(abs, body, "utf8");
+  const uri = vscode.Uri.file(abs);
+  const doc = await vscode.workspace.openTextDocument(uri);
+  await vscode.window.showTextDocument(doc);
+  await settle();
+  return { doc, key: uri.toString() };
+}
+
+// #40 (follow-up to #38): on undo/redo, restore each re-inserted char's EXACT
+// prior author attribution rather than leaving it neutral (#38). Driven mid-edit
+// (buffer dirty) so the attribution branch runs, not the disk-sync one.
+suite("F3 #40 — undo/redo restores exact author attribution (host E2E, no LLM)", () => {
+  test("undo of a deletion of AGENT text restores its agent span (not neutral, not human)", async () => {
+    const { doc } = await freshDoc("docs/s40agent.md", "Human start. xxxx end.\n");
+    const api = await getApi();
+    const ctl = api.attributionController;
+
+    // Make "xxxx" agent-authored via the seam → an agent span over "ROBOT".
+    const t = "xxxx";
+    const start = doc.getText().indexOf(t);
+    const ok = await ctl.applyAgentEdit(
+      doc,
+      new vscode.Range(doc.positionAt(start), doc.positionAt(start + t.length)),
+      "ROBOT",
+      AGENT,
+      { turnId: "turn-s40" },
+    );
+    assert.strictEqual(ok, true, "seam edit applies");
+    await settle();
+    const key = api.proposalController.keyFor(doc);
+    assert.ok(
+      ctl.getSpans(key).some((s) => s.authorKind === "agent" && doc.getText().slice(s.range.start, s.range.end).includes("ROBOT")),
+      "ROBOT is agent-attributed after the seam edit",
+    );
+
+    // Human deletes "ROBOT" (buffer stays dirty: it already diverged from disk).
+    const rs = doc.getText().indexOf("ROBOT");
+    const del = new vscode.WorkspaceEdit();
+    del.delete(doc.uri, new vscode.Range(doc.positionAt(rs), doc.positionAt(rs + "ROBOT".length)));
+    assert.ok(await vscode.workspace.applyEdit(del), "delete applied");
+    await settle();
+    assert.ok(!doc.getText().includes("ROBOT"), "ROBOT deleted");
+
+    // Undo the deletion → ROBOT re-inserted. #40: its AGENT span is restored.
+    await vscode.window.showTextDocument(doc);
+    await vscode.commands.executeCommand("undo");
+    await settle();
+    assert.ok(doc.getText().includes("ROBOT"), "undo restored ROBOT");
+    assert.ok(doc.isDirty, "buffer still dirty → attribution branch ran");
+
+    const r2 = doc.getText().indexOf("ROBOT");
+    const spans = ctl.getSpans(key);
+    const over = spans.filter((s) => s.range.start < r2 + 5 && s.range.end > r2);
+    assert.ok(over.length > 0, "restored ROBOT carries a span");
+    assert.ok(over.every((s) => s.authorKind === "agent"), `restored ROBOT is agent-attributed, got ${JSON.stringify(over)}`);
+  });
+
+  test("edit → undo → redo round-trips attribution to identical state", async () => {
+    const { doc } = await freshDoc("docs/s40roundtrip.md", "Base alpha. yyyy omega.\n");
+    const api = await getApi();
+    const ctl = api.attributionController;
+    const key = api.proposalController.keyFor(doc);
+
+    const start = doc.getText().indexOf("yyyy");
+    await ctl.applyAgentEdit(
+      doc,
+      new vscode.Range(doc.positionAt(start), doc.positionAt(start + 4)),
+      "BLUEWORD",
+      AGENT,
+      { turnId: "turn-s40b" },
+    );
+    await settle();
+    const norm = (k: string) =>
+      ctl
+        .getSpans(k)
+        .map((s) => ({ a: s.authorKind, t: doc.getText().slice(s.range.start, s.range.end) }))
+        .sort((x, y) => (x.t < y.t ? -1 : 1));
+    const afterAgent = JSON.stringify(norm(key));
+    assert.ok(afterAgent.includes("BLUEWORD") && afterAgent.includes("agent"), "agent span present after the edit");
+
+    // Delete BLUEWORD, then undo (restore) then redo (re-delete) then undo again.
+    const bs = doc.getText().indexOf("BLUEWORD");
+    const del = new vscode.WorkspaceEdit();
+    del.delete(doc.uri, new vscode.Range(doc.positionAt(bs), doc.positionAt(bs + "BLUEWORD".length)));
+    await vscode.workspace.applyEdit(del);
+    await settle();
+    await vscode.window.showTextDocument(doc);
+    await vscode.commands.executeCommand("undo"); // restore BLUEWORD
+    await settle();
+    assert.ok(doc.getText().includes("BLUEWORD"), "undo restored BLUEWORD");
+    assert.strictEqual(JSON.stringify(norm(key)), afterAgent, "attribution after undo matches the pre-deletion state exactly");
+  });
+});

test/e2e/suite/s48PinClean.test.ts

@@ -0,0 +1,63 @@
+import * as assert from "assert";
+import * as fs from "fs";
+import * as path from "path";
+import * as vscode from "vscode";
+import type { CowritingApi } from "../../../src/extension";
+
+const WS = process.env.E2E_WORKSPACE!;
+const settle = () => new Promise((r) => setTimeout(r, 400));
+
+async function getApi(): Promise<CowritingApi> {
+  const ext = vscode.extensions.getExtension("benstull.vscode-cowriting-plugin")!;
+  const api = (await ext.activate()) as CowritingApi;
+  assert.ok(api?.trackChangesPreviewController, "exports preview controller");
+  return api;
+}
+
+async function freshDoc(rel: string, body: string): Promise<{ doc: vscode.TextDocument; key: string }> {
+  const abs = path.join(WS, rel);
+  fs.mkdirSync(path.dirname(abs), { recursive: true });
+  fs.writeFileSync(abs, body, "utf8");
+  const uri = vscode.Uri.file(abs);
+  const doc = await vscode.workspace.openTextDocument(uri);
+  await vscode.window.showTextDocument(doc);
+  await settle();
+  return { doc, key: uri.toString() };
+}
+
+// #48 host E2E (no LLM): pinning the baseline leaves the review panel fully clean
+// — no authorship coloring on unchanged blocks — while re-divergence brings the
+// annotations back. The author colors are read from the on-state renderReview HTML.
+suite("S48 — pin → fully clean review panel (host E2E, no LLM)", () => {
+  test("pin clears authorship coloring; a later edit brings annotations back", async () => {
+    const { doc, key } = await freshDoc("docs/s48pin.md", "# S48\n\nAn original baseline paragraph.\n");
+    const api = await getApi();
+    const ctl = api.trackChangesPreviewController;
+    await vscode.commands.executeCommand("cowriting.showTrackChangesPreview");
+    await settle();
+
+    // Type a paragraph → a human attribution span + author coloring in the on-state.
+    const edit = new vscode.WorkspaceEdit();
+    edit.insert(doc.uri, doc.positionAt(doc.getText().length), "\n\nA freshly typed human paragraph.\n");
+    assert.ok(await vscode.workspace.applyEdit(edit), "operator edit applied");
+    await settle();
+    assert.match(ctl.renderHtmlFor(key), /cw-by-human/, "typed text is author-colored before the pin");
+
+    // Pin the baseline → zero diff → the panel must be fully clean.
+    ctl.receiveMessage(key, { type: "pinBaseline" });
+    await settle();
+    const pinned = ctl.renderHtmlFor(key);
+    assert.ok(!pinned.includes("cw-by-human"), "no human authorship coloring after pin");
+    assert.ok(!pinned.includes("cw-by-claude"), "no Claude authorship coloring after pin");
+    assert.ok(!pinned.includes("cw-add") && !pinned.includes("cw-del"), "no change marks after pin");
+    assert.match(pinned, /data-src-start/, "blocks still carry data-src offsets (INV-36 mapping kept)");
+    assert.ok(pinned.includes("freshly typed human paragraph"), "the body text is still rendered, just plain");
+
+    // Edit again → there are changes since the pinned baseline → annotations return.
+    const edit2 = new vscode.WorkspaceEdit();
+    edit2.insert(doc.uri, doc.positionAt(doc.getText().length), "\n\nA second typed paragraph diverges again.\n");
+    assert.ok(await vscode.workspace.applyEdit(edit2), "second operator edit applied");
+    await settle();
+    assert.match(ctl.renderHtmlFor(key), /cw-by-human/, "authorship coloring returns once the doc diverges from the pin");
+  });
+});

test/trackChangesModel.test.ts

@@ -168,6 +168,45 @@ describe("colorByAuthor", () => {
     expect(html).toContain('<span class="cw-by-human">hello</span>');
     expect(html).toContain("world");
   });
+
+  // #33: author-coloring must be sentinel-safe around markdown emphasis. These
+  // exercise the REAL markdown-it renderer (via renderReview on an unchanged doc),
+  // since the failure is in markdown-it's inline parsing / element nesting.
+  // CASE1 — a span boundary strictly inside a delimiter run must not split it.
+  test("#33 CASE1: a span boundary inside ** does not break emphasis parsing", () => {
+    const doc = "a**b**c";
+    const html = renderReview(doc, doc, [{ start: 0, end: 2, author: "human" }], []);
+    expect(html).toContain("<strong>b</strong>"); // emphasis still renders
+    expect(html).not.toContain("**"); // no raw delimiters left
+    expect(html).not.toContain("<em></em>"); // no stray empty emphasis (the parse-break symptom)
+    expect(html).toContain('class="cw-by-human"'); // coloring present
+  });
+  // CASE3 — a span boundary inside an emphasis run must color the text without
+  // misnesting span/element (the span is split at the element boundary).
+  test("#33 CASE3: a span boundary inside **bold** colors the text without misnesting", () => {
+    const doc = "**bold**";
+    const html = renderReview(doc, doc, [{ start: 0, end: 4, author: "human" }], []); // covers "**bo"
+    expect(html).toContain("<strong>");
+    expect(html).toContain('<span class="cw-by-human">bo</span>ld</strong>'); // span INSIDE strong, closed before "ld"
+    expect(html).not.toContain('cw-by-human"><strong>'); // NOT the old misnest (span wrapping the <strong> open)
+  });
+  // CASE2 — a span covering a whole emphasis run stays correct (regression).
+  test("#33 CASE2: a span over the whole **bold** colors it correctly", () => {
+    const doc = "**bold**";
+    const html = renderReview(doc, doc, [{ start: 0, end: 8, author: "human" }], []);
+    expect(html).toContain("<strong>");
+    expect((html.match(/cw-by-human/g) ?? []).length).toBe(1);
+    expect(html).toContain("bold");
+  });
+  test("#33: no Private-Use-Area sentinel chars leak into the rendered output", () => {
+    const doc = "a**b**c and `co de` and _x_";
+    const spans: AuthorSpan[] = [
+      { start: 0, end: 2, author: "human" },
+      { start: 12, end: 16, author: "claude" },
+    ];
+    const html = renderReview(doc, doc, spans, []);
+    expect(html).not.toMatch(/[\uE000-\uF8FF]/); // no leftover BMP Private-Use-Area sentinels
+  });
 });
 
 import { renderPlain } from "../src/trackChangesModel";
@@ -277,6 +316,54 @@ describe("renderReview", () => {
     expect(uIdx).toBeGreaterThan(html.indexOf("Beta there")); // unanchored still trails the body
     expect(html).toContain("cw-proposal-unanchored");
   });
+  // #48: a PINNED baseline with zero changes leaves the panel fully un-annotated —
+  // no authorship coloring on unchanged blocks — while pending proposals still show.
+  test("renderReview: pinned + zero diff with author spans renders NO authorship coloring", () => {
+    const doc = "Human wrote this.\n\nClaude wrote that.";
+    const spans: AuthorSpan[] = [
+      { start: 0, end: 17, author: "human" },
+      { start: 19, end: doc.length, author: "claude" },
+    ];
+    const html = renderReview(doc, doc, spans, [], { pinned: true });
+    expect(html).not.toContain("cw-by-human");
+    expect(html).not.toContain("cw-by-claude");
+    expect(html).not.toContain("cw-add");
+    expect(html).not.toContain("cw-del");
+    // still a selection→source surface (INV-36): blocks carry data-src offsets.
+    expect(html).toContain("data-src-start");
+    // the body text is still there, just plain.
+    expect(html).toContain("Human wrote this.");
+    expect(html).toContain("Claude wrote that.");
+  });
+  test("renderReview: pinned + zero diff still renders a pending proposal block", () => {
+    const doc = "Human wrote this.\n\nClaude wrote that.";
+    const spans: AuthorSpan[] = [{ start: 0, end: 17, author: "human" }];
+    const proposals: ProposalView[] = [
+      { id: "p1", anchorStart: 0, anchorEnd: 5, replaced: "Human", replacement: "Person" },
+    ];
+    const html = renderReview(doc, doc, spans, proposals, { pinned: true });
+    expect(html).not.toContain("cw-by-human"); // body still clean
+    expect(html).toContain('data-proposal-id="p1"'); // proposal still shows (it is an action)
+    expect(html).toContain("Person");
+  });
+  test("renderReview: zero diff WITHOUT a pin (e.g. machine-landing) keeps authorship coloring (INV-33)", () => {
+    // accepting a Claude edit advances the baseline (zero diff) but is NOT a pin —
+    // the landed author coloring must remain.
+    const doc = "Human wrote this.\n\nClaude wrote that.";
+    const spans: AuthorSpan[] = [{ start: 19, end: doc.length, author: "claude" }];
+    const html = renderReview(doc, doc, spans, []); // no pinned flag
+    expect(html).toContain("cw-by-claude");
+  });
+  test("renderReview: with REAL changes since a pin, author coloring returns", () => {
+    // a genuine added block is still author-colored even when pinned (only the
+    // zero-diff-after-pin state is clean).
+    const baseline = "Hello world";
+    const current = "Hello world\n\nHello world";
+    const spans: AuthorSpan[] = [{ start: 19, end: 24, author: "human" }];
+    const html = renderReview(baseline, current, spans, [], { pinned: true });
+    expect((html.match(/cw-by-human/g) ?? []).length).toBe(1);
+  });
+
   test("renderReview is deterministic with mixed anchored/unanchored proposals", () => {
     const doc = "one two three";
     const proposals: ProposalView[] = [