test(editor): golden round-trip conformance suite across both serializers

[h4yfans]

Closes #1848. Part of epic #1843 — this is the suite that makes the whole token class fail in CI instead of in a user's inbox.

Shape

One corpus, @memry/editor-schema/conformance, asserted byte-for-byte by both pipelines:

• Renderer half — roundtrip-conformance.test.ts (renderer suite): a REAL BlockNoteEditor on the real editor schema, through parseMarkdownPreservingBlanks → normalizeNoteBlocks → serializeBlocksPreservingBlanks — the actual open→save cycle, real remark escaping included. First unmocked renderer round-trip coverage.
• Main half — blocknote-converter.roundtrip.test.ts (main suite): markdownToYFragment → yDocToMarkdown, extending the assertions already in blocknote-converter.test.ts.

Each case asserts (1) round-trip identity against canonical bytes, (2) second-pass idempotence (write-back never rewrites what it just wrote). Both halves target the SAME canonical bytes, so cross-serializer byte equality is a tested equality, not an assumption.

Coverage

Family	Cases
Link mention URLs	_ * ! ~ ' ( ) % & = #, spaces, non-ASCII, trailing ), kitchen-sink, underscore-next-to-punctuation
Date pills	body text, crafted base64url __/-- emphasis runs, reminder variants, two-in-a-line
Callouts	all four types, multi-line body, foreign > [!note]/> [!tip] passthrough, title-after-marker, multi-paragraph body, nested foreign
Toggles	empty, body, nested, blank lines, code fence, image, unterminated
Containers	mention+date in table cells / list items / toggle bodies, wiki link + hash tag in cells and sentences, callout in toggle
Block markers	youtube embed, bookmark, file marker, task line
Golden vault	four new roundtrip-*.md fixtures, byte-gated by the existing frontmatter suite AND pushed through the converter

Fuzz

No property-testing dep in the workspace and pnpm add is off-limits, so a seeded mulberry32 drives 48 deterministic cases per family (mention URL alphabet, date payload alphabet incl. bytes that force base64url _/-, callout bodies, structured toggle bodies incl. nesting and fences). Same bytes every CI run. Free-text alphabets exclude emphasis pairs/backticks deliberately: remark canonicalizes those in ANY paragraph — accepted app-wide behavior, not this bug class. Inside a token nothing may be rewritten, so the token alphabets keep every hostile character.

Expected failures (the two-phase contract)

Cases broken on current main run as it.fails tagged with the sibling that fixes them — when that sibling lands, the inverted expectation turns red and forces the flag off, converting the case into a permanent assertion of the fixed behavior:

• Link mentions degrade to literal ((mention:…)) text after reload #1844 (link mentions): underscore-next-to-punctuation URL (both pipelines), mention-in-table-cell (renderer — rich render rewrites the token as a markdown link), mention URL fuzz family (both).
• Callouts reopen as a bare [!info] line above their content #1846 (callouts): foreign [!note]/[!tip] rewritten to [!info] (renderer), title-after-marker split (renderer), multi-paragraph body collapse (both), nested foreign callouts (both).
• Toggles spill their contents; fold state is never persisted #1847 (toggles): unterminated <details data-memry-toggle> currently DROPS its open/summary lines on both pipelines.
• Wiki link in a table cell serializes as bare display text on the renderer path #1865 (new, found by this suite): a [[wiki link]] in a table cell serializes as bare display text on the renderer path.

#1845 (date pills): every enumerated and fuzzed payload already round-trips on both pipelines — the suite locks that in as plain green assertions.

Telemetry

unclaimed-token-telemetry.ts, wired at the end of normalizeNoteBlocks (every note surface runs that chain): counts ((mention: / ((date: still literal after promotion and paragraph-leading orphaned [!type] markers, skipping code blocks. Reports through the existing app_error_seen channel as action: editor_unclaimed_token, errorCode: unclaimed_{mention,date,callout_marker}, count in metrics.itemCount. Leading emit on first sighting, then at most one aggregated event per kind per minute — the chain re-runs per note open/remote update and the main-side error path is itself per-minute throttled, so the counter aggregates instead of spamming. Metric only, no toast. Documented in architecture/observability.md.

CI note

check:architecture (inside pnpm typecheck) is red on current main for every branch (apps/mobile/vitest.config.ts -> node:url); fix is #1868. This branch stays pure of that fix and goes green once #1868 merges and it rebases.

Verification

• main suite: 48 passed, 5 expected-fail · renderer suite: 38 passed, 10 expected-fail · telemetry: 7 passed
• pnpm lint 0 errors; typecheck:node/web/test green (new test files compile, exclude backlog untouched); docs:impact --strict covered; docs:build green; full main project: only the pre-existing check-cert-hashes-config local-env failures (identical on untouched main); full renderer project: failures identical to the pre-existing set on main (baseline diff run to confirm).

[github-actions]

React Doctor found no new issues. 🎉

_{Reviewed by React Doctor for commit 7e28137.}

[codecov]

Codecov Report

❌ Patch coverage is 98.73418% with 1 line in your changes missing coverage. Please review.

Files with missing lines	Patch %	Lines
...nts/note/content-area/unclaimed-token-telemetry.ts	98.68%	1 Missing ⚠️

📢 Thoughts on this report? Let us know!