bug(omp): first tag revision allocation in a nonempty legacy session invalidates its snapshot #769

What happened

I resumed an older OMP session and my next prompt was refused:

Magic Context could not safely prepare this turn; send your message again.
stage=pipeline elapsed=572ms recovery=no managed result; refused
(Pi tag hydration: tag snapshot is not ready; turn not sent, retry)

Hydration initially succeeded. The log then showed 25 pending operations applied, followed by a failure in runPipeline → getPiTagSnapshotView → sync.

Reproduction and likely cause

I reproduced the same error using the installed snapshot reader and a disposable database copy:

1. Keep an existing session's tags, but remove its schema_migrations_meta key tag_revision:<sessionId> to simulate legacy state.
2. Call ensurePiTagSnapshotReady, then read the snapshot successfully.
3. Update a tag's status on the same connection.
4. Read the snapshot again: tag snapshot is not ready.

With the revision key present, the same sequence passes. The first local update allocates a revision incarnation, changing foreignRevision. The existing firstAllocationPending handling only covers an empty session, so this nonempty legacy case is rejected.

I do not have the database from immediately before the original failure, so this establishes a reproducible mechanism, not proof of every original precondition.

Current status / expected behavior

The affected session now has a revision record. Offline replay of its current state passes context preparation (1,229 → 941 messages) and the OMP provider structure check. I have not verified a live model response.

I expect a session's own first revision allocation to preserve snapshot validity, while actual foreign writes and recreated revisions must still invalidate it. Please add regression coverage for nonempty legacy sessions and retain the foreign-write checks.

Environment

• Magic Context 0.49.16; OMP 18.12.6; macOS arm64.
• Offline reproduction: Bun 1.4.2.
• The local OMP replay mitigation described in bug(omp): tag stripping leaves unsigned text with stale Responses payload and blocks subsequent turns #766 was installed, but matched zero messages in this session. This failure occurs earlier, in the pipeline.
• Pi/OpenCode 1/OpenCode 2 were not tested for this failure.
• Distinct from the worker deadline/ack deadlock in Tag snapshot hydration worker deadlocks under Bun (credit ack never delivered after first transferred packet) #754.