After leaving Rust mode on OpenCode 1, the TS historian refuses forever: "gap before message N" on synthetic-only user rows the module skipped #781

Versions: plugin 0.49.22 on OpenCode 1. The code paths cited below are unchanged on main (3a6d3cb90).

What happens: a session ran in Rust mode for weeks. During that time a plugin injected synthetic: true user text messages into it (a cache-warm ping, one text part, about 56 chars). The module's compartments do not cover those rows: each one sits alone between two compartments.

After the project switched back to TS mode, every historian run fails before any model call:

historian failure: source=existing-validation reason="gap before message 3387 (expected 3386)"

It failed 66 times in 8 hours, and the session got no new compartments. The stored compartments had four such one-message gaps (3386, 3770, 4180, 4220). Each gap message was a user row whose only part was a text part with synthetic: true.

Why it never recovers (main):

• validateStoredCompartments (compartment-runner-validation.ts:302-321) accepts a gap only if it is listed in nonNarrativeGapRanges.
• v2NonNarrativeStoredGapRanges (compartment-runner-incremental.ts:143) returns [] unless the session's rebase generation is v2, so on OpenCode 1 no gap is ever accepted.
• isStrictGapHealingMessage (read-session-raw.ts:20) would also reject these rows: a V1 row has no storeType: "synthetic", and a text part with content returns false, even with synthetic: true on it.
• The repair and backoff around existingValidationError (compartment-runner-incremental.ts:394-430) apply only when an unresolved compartment is involved. This failure therefore repeats on every trigger and has no repair path.

Workaround (the same shape as the one confirmed on #561): for each gap, set the next compartment's start_message back by one, and its start_message_id to the gap message. The next run published normally.

Suggested fix:

1. Treat a V1 user row whose parts are all synthetic: true text as gap-healing, as storeType: "synthetic" already is on V2.
2. Accept such gaps on OpenCode 1 too, not only after a v2 rebase.
3. Alternatively, heal the gap by extending the next compartment, as the workaround does.
4. Either way, back off this failure class like the unresolved case, so it doesn't re-run on every trigger.

Related: #561, where a compaction marker on a synthetic user row wedged a session in Rust mode.