fix(store): prevent FileStore persistence starvation during continuous writes

[DeliciousBuding]

Summary

FileStore's background loop resets its 50 ms debounce timer for every write notification. A sustained stream with no 50 ms quiet period can therefore postpone persistence indefinitely, including the earliest accepted message. This is a correctness/liveness issue in the supported --store-backend file and legacy --store-file paths, not a claim about production request frequency or JSON throughput.

On source base 81e02d403a48fdb5130c233ed2da9b8919d2a3fb, a regression using the real NewFile constructor and CreateProject/CreateThread/CreateItem methods feeds 40 further writes at half the existing interval. After 20 batch intervals, memory has 41 items and the actual disk restore path has zero. No final Flush/Close is allowed to repair the file before inspection. testing/synctest drives virtual scheduling time; the snapshot and restore I/O are real.

Observed failure: continuous writes starved FileStore persistence for 20 batch intervals: memory items=41, durable items=0.

The earlier #2256 / #2333 / #2334 work measured explicit full-snapshot flush cost and retained that storage format. It did not exercise background notification timing. This issue preserves full-snapshot persistence and does not reopen those throughput candidates. Assistant transcript emission is buffered until run completion; this is not a per-token transcript-write claim.

Scope / file ownership

One writer in an isolated task worktree. Allowed files:

• edge-server/internal/store/file_store.go: arm the existing 50 ms timer on the first pending write; do not let later writes push the batch deadline forward; re-arm for later batches and stop the timer on shutdown.
• edge-server/internal/store/file_store_persist_loop_test.go: deterministic real-file constructor-loop tests for sustained progress, coalescing, new batches, idle behavior and pending-close persistence.
• docs/architecture/02-edge-server.md: describe bounded batching rather than trailing-edge-only debounce.

Interfaces / invariants / exclusions

No public API, snapshot schema, full-snapshot encoder, Sync/rename sequence, channel capacity, backend defaults or runtime settings change. Writes remain asynchronous/non-blocking with respect to disk persistence. The existing 50 ms value becomes a first-write batch scheduling window rather than an indefinitely extensible quiet period; it is not an fsync latency or durability-under-I/O-failure SLA. No idle polling writes and no per-notification synchronous flush. Preserve Flush, Close, LastPersistError and post-close memory-write behavior; do not add automatic retry policy or a new storage abstraction. No deployment, restart, production load test or real process-kill claim.

Acceptance

• Reproduce durable starvation on unchanged production source using real file I/O and the constructor-owned loop.
• Sustained writes make repeated durable progress before the stream becomes idle, without final Flush/Close.
• Short bursts coalesce until the first-write deadline; later writes cannot postpone it; later batches still persist.
• An idle store does not persist repeatedly and existing error reporting/close flushing behavior remains effective.
• Complete relevant Windows and Linux race tests, vet/staticcheck and document checks.
• All required PR checks pass on the final head.
• Merge, synchronize source-only owner notes and clean the completed branch/worktree.

[DeliciousBuding]

Verified closeout

#2346 merged as a3f044b19189ff7932b7a5b67c0a13f9baeeaf81. The merge and tested head 10f40b20205633838d050b70c599bea7bf1a5a8e share tree a223635e6c3bb5bf42e5291a1b4a4d23e09ebe8b.

• Fixed continuous-write starvation while preserving fixed-window coalescing, idle behavior, explicit error reporting and close flushing. Four real-file/synctest regressions and four behavioral negative controls verified the contract; no final Flush/Close concealed the initial starvation failure.
• Store/API/CLI Windows and Linux race tests, vet/staticcheck and existing document/contract checks passed. An additional complete Windows Edge short suite passed with real package output (29.845 s, exit 0).
• CI run 34036655934 attempt 2 passed all seven required checks on the unchanged head. Attempt 1 passed package tests but failed deleting lifecycle.test.exe with Access is denied; its exact lock owner remains unknown. The failed job/dependents were retried after diagnosis, not hidden by weakening tests. CodeRabbit reported Review skipped, not approval.
• Stable architecture documentation and source-only operational notes are synchronized and committed. The completed FileStore task branch/worktree has been removed; unrelated work and original evidence are preserved. New delivery-admission work is tracked separately in fix(dispatch): make Hub-Edge task admission retry-safe end to end #2347.

No runtime rollout, restart, retained-data/schema change, real-model request or packaged Desktop claim. #2304 and #1663 remain independent follow-ups.