Warn when persisted segment boundaries are lost on load (reversification / source edits)

[alex-rawlings-yyc]

Problem

Segment boundaries are persisted as a SegmentationDelta anchored to token refs of the form "<verse SID>:<charStart>" (e.g. "PSA 3:2:14"). On load, resegmentBook → effectiveStarts/normalize (src/utils/segmentation.ts) silently drop any anchor whose ref isn't found in the freshly-tokenized book, reverting that region to the default one-segment-per-verse segmentation.

This is the intended, non-destructive degradation — but it's silent. When the source text is reversified (classically the Psalms verse-0↔verse-1 superscription shift, which renumbers SIDs) or edited upstream (shifting char offsets), a user's hand-placed boundaries vanish with no signal. They lose work and may not notice.

Decisions made by Alex (up for debate)

Detect-and-warn — not remap. Reframed as a general "lost boundaries" safety net (reversification is just one trigger; ordinary edits are the same event to an anchor):

• Keep the existing silent drop (non-destructive). Add a warning.
• Detection signal: any non-empty SegmentationDelta that loses anchors on load. Do not try to isolate reversification from ordinary edits — this deliberately sidesteps the host's versificationStr, which is documented as unreliable (InterlinearNavContext.tsx:51). A project with default/empty segmentation never warns.
• Do NOT prune the dead anchors from storage. Keeping them preserves the model's revival property: if the source reverts and the token ref reappears, the boundary comes back on its own.
• Explicitly out of scope: remapping boundaries across versifications, and any versification-mapping API. That's a much larger effort with poor cost/benefit for a rare event; not this issue.

Implementation notes (from a prior investigation)

• Set-difference site: the book memo in src/components/InterlinearizerLoader.tsx (~line 272) is where the freshly-tokenized verseBook and the delta (draft?.segmentation) are both in scope. Detection must run in a useEffect, not the pure book memo (notifications are a side effect).
• Add a new pure helper (e.g. lostAnchors(verseBook, delta): string[]) that intersects [...removedVerseStarts, ...addedStarts] against the book's token refs and returns the missing ones. Do not make effectiveStarts/normalize report drops — keep segmentation.ts pure. (Note: bookLookups in that file already computes an all token-ref set you can mirror.)
• Delivery: papi.notifications.send({ message, severity: 'warning' }) (PlatformNotification; Severity = 'info' | 'warning' | 'error'). No frontend code sends notifications yet, but the mock already exposes papi.notifications.send.
• Copy: generic — "source text changed," not "versification changed" (we can't reliably attribute the cause). Optionally include a count. Add a key to contributions/localizedStrings.json, resolve via useLocalizedStrings, interpolate the count with .replace('{count}', () => …) (pattern in MorphemeBox.tsx / TokenChip.tsx).
• Dedup: the extension re-tokenizes on many events (duplicate GetText re-fetches), so a naive per-load warning would spam. useInterlinearizerBookData returns a reference-stable verseBook across duplicate USJ payloads (it dedupes by JSON), making the verseBook reference a natural dedup anchor (track the last-warned one in a useRef).

⚠️ Open decision — resolve before implementing

Dedup semantics were not settled:

• (A) Per-tokenization: warn again on each genuinely new re-tokenization that still has lost anchors; duplicate re-fetches suppressed via the reference-stable verseBook. (Better matches "warn until something meaningful changes"; recommended in the prior investigation.)
• (B) Strict once-per-session: a projectId+book Set that never re-warns for that book again this session — even across a reversify → revert → reversify cycle.

Pick one before building.

Acceptance / tests

1. A lossy non-empty delta → warning sent (positive).
2. A delta whose anchors all still resolve → no warning (negative).
3. Default/empty segmentation → no warning (negative).
4. Repeat render with the same verseBook → no repeat warning (dedup).
5. Dead anchors are not pruned — the persisted/draft segmentation is untouched by the warning path.

100% coverage enforced; npm run lint-fix at the end.
`