Summary

Every 5s-debounced session delta sync calls sync({ reason: "session-delta" }) with no targets, so syncSessionFiles performs a full sessions-directory enumeration — and then skips every listed file that is not in the dirty set anyway. The enumeration's only unique contribution to a delta sync is stale-row prune reconciliation; the indexing work is fully determined by sessionsDirtyFiles, which the engine populated from transcript events before triggering the sync.

On networked filesystems these scans are expensive: an NFSv4.1 profile of a real deployment measured full listings of agents/main/sessions at 3.4–3.8s each (~100 entries; the client requests a broad READDIR attribute set, which the backing server satisfies with per-entry metadata reads). Trace-level attribution later showed the specific scans that profile captured came from per-save archive-retention cleanup — fixed separately in #91957 — not from memory delta syncs (that deployment does not run memory session sync). The per-scan cost numbers still stand, and they are what every delta sync pays on such storage whenever memory session sync is enabled.

This PR makes delta syncs use the targeted path that already exists (runMemoryTargetedSessionSync):

• processSessionDeltaBatch now passes the accumulated dirty set as explicit sessionFiles targets → the sessions-dir enumeration is skipped entirely. The targeted path clears only the synced files from the dirty set (files dirtied mid-sync survive) and retains targets as dirty on failure.
• Bounded reconciliation: at most once per SESSION_PRUNE_RECONCILE_INTERVAL_MS (15 min), a delta sync falls back to the full enumeration so index rows for files deleted out-of-band (another process/node on shared storage) still get pruned. The timestamp only advances when an authoritative prune actually completed, so a failed scan retries on the next batch.
• Boot catch-up, interval-with-dirty-files, CLI/forced, and full-reindex syncs keep their full enumerations unchanged — those are the intended reconciliation points.

Behavior notes

• Indexing outcomes are unchanged. The full path already skipped non-dirty files during delta syncs; only the prune cadence shifts from per-batch (~5s) to the bounded window (≤15 min during activity). Idle agents are unchanged (no delta syncs → no scans, before and after).
• Targeted syncs are sessions-only by existing contract, so a delta sync no longer incidentally flushes dirty memory-source files; those still sync via their own watch debounce and interval triggers.
• Stacked on fix(memory): do not prune session index from a failed directory scan #91091 deliberately: the now-rarer full enumerations are exactly the scans fix(memory): do not prune session index from a failed directory scan #91091 makes authoritative ({ ok, files }), so a failed reconcile scan cannot prune and dirty state is retained for retry.

Identified while decomposing the NFS readdir-load report behind the #91081 discussion (its "avoid full scans when a targeted dirty-file set exists" recommendation). With this change, deployments running memory session sync stop paying a full directory scan per delta batch; the remaining enumeration cost (boot catch-up, periodic reconcile, qmd export, and the store-save retention sweeps addressed by #91957) is the much smaller surface the coalescing/TTL discussion in #91081 addresses separately.

Verification

• node scripts/run-vitest.mjs extensions/memory-core/src/memory/manager-sync-ops.targeted-delta.test.ts extensions/memory-core/src/memory/manager-sync-ops.prune-guard.test.ts extensions/memory-core/src/memory/manager-targeted-sync.test.ts extensions/memory-core/src/memory/manager-session-sync-state.test.ts — 19/19 (5 new).
• node scripts/run-vitest.mjs extensions/memory-core/src/memory/manager-sync-ops.archive-delta-bypass.test.ts — 4/4.
• pnpm tsgo:extensions and pnpm tsgo:extensions:test — clean. node scripts/run-oxlint.mjs on touched files — clean.
• New tests cover: targeted params within the reconcile window; accumulated dirty set (leftovers from earlier syncs) included in targets; full-enumeration fallback at window expiry (and at boot, timestamp 0); reconcile timestamp advancing only after an authoritative prune; failed scan leaving the timestamp unset so the next batch retries.
• Not run here: live NFS before/after trace. The report's verification protocol applies directly — count scandir events on the sessions dir during steady-state activity; expected: one per 15-min reconcile window instead of one per 5s delta batch.