[Bug]: Durable memory maintenance rescans per search, forces full reindex on chunking bump, and replays provenance backfill on every open

Bug type

Behavior bug (incorrect output/state without crash)

Beta release blocker

No

Summary

Several defects in durable-memory maintenance:

1. openclaw memory search and the memory.search gateway RPC build a transient manager per call — every search rescans and hashes the whole corpus even though the gateway already owns a watched, current index; recall signals from operator searches are dropped (never recorded).
2. The chunking v2→v3 bump forces a full semantic reindex even for Markdown-only memory roots whose chunk boundaries are identical across versions.
3. ensureMemoryChunkProvenance takes an immediate write lock and replays the whole backfill on every manager open, even when provenance is already complete.
4. Promotion candidates with blocked origins are ranked and narrated even though the durable write boundary always rejects them.
5. memory-wiki's bridge import has no path exclusion mechanism.

(The related memory-flush append-duplication defect is tracked separately by PR #120404.)

Steps to reproduce

1. With a populated memory root and a running gateway, run openclaw memory search <query> repeatedly; observe a full corpus rescan/hash per invocation instead of the gateway's live index being used.
2. Upgrade across the chunking v2→v3 boundary with a Markdown-only root; observe a full re-embedding despite identical chunking.
3. Open the manager repeatedly; observe the provenance backfill write path taken each time.

Expected behavior

Searches reuse the gateway's live manager and record recall; version bumps that provably do not change chunk boundaries upgrade metadata in place; provenance ensure short-circuits when complete; blocked-origin candidates are excluded from ranking; bridge imports support exclusions.

Actual behavior

Per-search rescans with dropped recall signals; unnecessary full reindexes; repeated backfill lock/replay; dead-end promotion narration.

OpenClaw version

2026.8.1 (source checkout)

Operating system

Ubuntu 26.04 LTS (Linux 7.0.6-2-pve x86_64)

Install method

Local pnpm source build

Model

N/A (embedding provider-independent)

Provider / routing chain

N/A

Additional provider/model setup details

N/A

Logs

NOT_ENOUGH_INFO

Screenshots, recordings, and evidence

Traceable in source: transient manager construction in the memory search CLI/RPC paths; unconditional reindex trigger on chunking version mismatch; unconditional provenance backfill in packages/memory-host-sdk.

Impact and severity

Slow searches, lost recall telemetry, and expensive unnecessary reindexes for every operator with a large memory corpus. A fix exists on a local branch (route-first gateway-backed search with --local fallback and recall recording, Markdown-only in-place chunking upgrade, provenance short-circuit, batched embedding-cache seeding with a query LRU, wiki bridge exclusions); PR to follow.