Fix non-Latin record titles collapsing to identical untitled slug

[grimicorn-agent]

What & why

titleToSlug in server/utils/markdown.ts ran .toLowerCase().replace(/[^a-z0-9\s-]/g, ""), which deleted every non-ASCII letter. Any title written in a non-Latin script (Cyrillic, CJK, Arabic) therefore produced an identical empty base and collapsed onto untitled, forcing untitled-1/untitled-2 disambiguation, and accented Latin titles were mangled (Café → caf).

This normalizes the title with Unicode NFKD and drops combining marks before slugifying:

• Accented Latin folds to its ASCII base: Café Meeting → cafe-meeting, Zürich Naïve → zurich-naive.
• Non-Latin scripts survive NFKD with no ASCII, so they strip cleanly to the FALLBACK_SLUG untitled — the fallback is applied deliberately and downstream resolveUniqueFilePath handles uniqueness (per the issue's acceptance criteria).
• Non-ASCII whitespace (NBSP U+00A0, ideographic space U+3000) NFKD-normalizes to a plain space, so it still acts as a word separator instead of fusing words.

Decisions

• NFKD normalization over transliteration. The acceptance criteria call for non-Latin titles to yield FALLBACK_SLUG, not for transliterated slugs. NFKD achieves that while also improving the accented-Latin case, with no new dependency.
• Operation order: normalize → strip marks → lowercase → strip. Lowercasing first would delete ASCII emitted by compatibility decomposition of uncased symbols (№ → No → o). Normalizing first preserves it (№5 Meeting → no5-meeting).
• Named regex constants (COMBINING_MARKS_PATTERN, NON_SLUG_CHAR_PATTERN) replace the inline literal for clarity.

Tests

Added titleToSlug cases in tests/server/utils/markdown.test.ts: non-Latin-script titles → untitled, accented-Latin folding, compatibility-decomposition ASCII survival, and non-ASCII whitespace as a separator. Full suite: 1753 passing. lint:fix, lint:ci clean.

Viewable

Server-side slug utility; exercised on ingest via parseWebhookPayload / parseEmailPayload (webhook + email routes). Behavior is covered by tests/server/utils/markdown.test.ts.

Closes #229

Follow-up suggestions

• Fold stroke/ligature Latin letters — letters with no NFKD decomposition (ß, ø, ł, æ, þ, đ) are stripped rather than folded, so Straße → strae; an explicit fold map would produce readable slugs (suggested: P3, effort: S, evidence: server/utils/markdown.ts titleToSlug)
• Bound title length before NFKD normalization — NFKD can expand a string up to ~18x (e.g. U+FDFA), and titleToSlug normalizes the full untruncated title; pre-slicing would cap the transient allocation (suggested: P4, effort: S, evidence: server/utils/markdown.ts titleToSlug)

[grimicorn-agent]

Independent code review trail

Reviewer: fresh Claude (Opus), did not write the code. 3 rounds.

Round 1 — flagged:

• Explicit non-ASCII strip was a functional no-op vs. the existing whitelist, AND a regression: it deleted non-ASCII whitespace (NBSP, ideographic space) that previously became hyphens. Fixed — replaced the strip with NFKD normalization + combining-mark drop, which folds accented Latin (café → cafe), normalizes non-ASCII whitespace to a plain space, and still strips non-Latin scripts to untitled.
• Tests didn't discriminate the change / comment overclaimed. Fixed — added café→cafe-meeting (fails on the old regex, verified) and reworded the comment.

Round 2 — flagged:

• toLowerCase() ran before normalize("NFKD"), so ASCII emitted by compatibility decomposition of uncased symbols was deleted (№ → No → o). Fixed — reordered to normalize → strip marks → lowercase → strip; added №5 Meeting → no5-meeting test.
• titleToSlug("Привет") === titleToSlug("こんにちは") asserts nothing (passes on a constant return) and is redundant. Fixed — removed; the assertion above already pins each non-Latin title to untitled with concrete values.

Round 3 — flagged, all declined with reasons:

• Add a fold map for stroke/ligature Latin (ß, ø, ł…). Skipped — pre-existing (the old code mangled these identically; not a regression), and the acceptance chose strip/normalize over transliteration. Recorded as a follow-up suggestion.
• Replace stripped chars with a space instead of deleting. Skipped — changes existing ASCII behavior (don't → don-t instead of dont), out of scope.
• Hash-suffix the untitled fallback so distinct non-Latin titles get distinct filenames. Skipped — directly contradicts the acceptance ("yields the FALLBACK_SLUG 'untitled'"); downstream resolveUniqueFilePath already disambiguates collisions.
• Bound title length before NFKD. Skipped — low severity, bounded by request body limits; a correct bound needs a non-obvious multiplier. Recorded as a follow-up suggestion.

Unresolved after 3 rounds: none blocking. Two declined items captured as follow-up suggestions in the PR body.

[grimicorn-agent]

Merge maintenance + review trail

Merged origin/main into agent/nonlatin-slug (was BEHIND). No conflicts — main's changes (new server routes, filenameTemplate.ts, vault utils, etc.) don't overlap this PR's server/utils/markdown.ts change. After merge: npm run lint:fix clean, npm run lint:ci clean, npm run test green (1842 passed / 119 files).

Independent review (Opus)

Round 1 flagged code-standards clean (small function, full names, no nesting). Behavioral findings, all skipped with reasons:

• Non-Latin titles collapse to untitled — this is the PR's deliberate, code-commented design (transliteration folds accented Latin; non-Latin scripts fall back to FALLBACK_SLUG). Collision risk is mitigated upstream: buildFilename composes the slug with {{date}} and {{source}} template tokens (documented in shared/utils/filenameTemplate.ts as the collision guard). Reviewer confirmed it is not introduced by this diff. Redesigning the fallback (e.g. appending an id) is out of scope for a behind-main merge and titleToSlug has no id in scope. Skipped.
• Adjacent non-Latin/ASCII merges words (Deploy部署v2 → deployv2) — edge case of the intended strip-to-empty behavior; changing it to strip-to-space is a design choice belonging to the PR author, not a merge task. Skipped.

No correctness/security/standards issues to fix in the diff, so no code changes were made and no further rounds were needed.