fix(markdown): persist browser editor changes [5/5]

[George Ng (GeorgeNgMsft)]

Summary

Final layer of the replacement stack for #2940, rebased onto the latest #2969 (georgengmsft-markdown-browser-server).

• persist browser autosave and explicit saves with binding tokens and expected revisions
• preserve serialized Markdown syntax and surface serializer failures
• reject stale or incomplete binding, snapshot, switch, and revision data
• reconcile identical-content conflicts while retrying transient binding contention
• preserve nested document paths, opaque collaboration room IDs, and browser history
• coordinate primary autosave ownership without allowing secondary or reconnect events to bless unseen content

Adversarial review

Two independent principal-level reviews used GPT-6 Astra and Grok 4.6. Their validated findings were fixed:

• stale content could overwrite newer disk content after primary promotion, secondary autosave events, or SSE reconnects
• temporary update-lock conflicts could permanently suppress autosave retries
• initial editor content could be treated as a user edit and rewritten on the first autosave interval
• Back/Forward navigation could add history entries or resolve rapid switches to the wrong document
• memory-only secondary clients could fail to become the primary autosave client after election
• accepted snapshots could retain stale conflict suppression

Focused regression coverage was added for these cases. No unresolved high-confidence findings remain after synthesis and follow-up review.

Validation

• pnpm run build packages/agents/markdown
• full Markdown local suite: 12 suites, 113 tests passed
• pnpm run prettier:changed
• lint ratchet: passed (changed-file violations 182 → 178)
• complexity ratchet: passed (over-threshold counts unchanged; no new-file function over limits)
• circular-dependency ratchet: passed (220 → 220)
• test-debt gate: passed (no focused or newly skipped tests)
• git diff --check

The local dependency preflight was run in warning mode because this worktree's existing native better-sqlite3 rebuild state caused automatic pnpm install to fail before commands started. The actual TypeScript/Vite build, Jest tests, formatting, and all CI ratchets completed successfully against the installed dependency graph.

The rebased layer changes 1,216 lines across 9 files (+1,068/-148), below the 1,500-line cap.