Summary

Moves the maintenance-critical part of LCM off transcript-file ownership by treating OpenClaw's canonical SQLite transcript scope as the source of truth when it is available.

This PR closes the two real downstream survivors we were still seeing after the earlier slices:

• transcript GC candidate matching still depended on reopening the transcript file even when SQLite was canonical
• session_end(reason=\"compaction\") still split continuity into a fresh segment instead of preserving one logical conversation

This is the implementation slice for #644, stacked on top of #643.

Why this is better

Once OpenClaw's database-first runtime is canonical, LCM should behave like a companion projection system, not like the raw transcript owner.

That means:

• maintenance should target canonical transcript scope, not file possession
• checkpoint refresh should follow the canonical frontier, not only file stats
• compaction successor runtime sessions should preserve one logical segment unless OpenClaw actually signals a real lifecycle split

This PR makes those boundaries real.

Architecture

flowchart LR
    A["OpenClaw canonical SQLite transcript scope"] --> B["LCM maintenance / GC mapping"]
    B --> C["summary-store GC candidates"]
    B --> D["rewrite request to host runtime"]
    A --> E["frontier refresh"]
    F["session_end reason=compaction"] --> G["same family, same segment"]

What changed

Canonical transcript GC mapping

• adds SQLite-aware tool-result entry-id mapping for transcript GC candidate replacement
• stops requiring transcript-file ownership when the host exposes canonical SQLite transcript replay
• refreshes bootstrap checkpoint state from the SQLite frontier after maintenance when SQLite is the canonical source

Continuity correction for compaction successors

• treats session_end(reason=\"compaction\") as successor continuity instead of archive-and-replace
• preserves the same conversation family and segment while updating the runtime sessionId

Regression coverage

• proves maintenance can target canonical SQLite transcript scope without reading a transcript file
• proves successor runtime sessions can resume SQLite bootstrap inside the same logical segment

Non-goals

This PR does not:

• redesign the upstream OpenClaw typed projection seam from #79905
• move LCM summaries or recall state into OpenClaw core
• remove legacy JSONL-only manual rotation paths for older hosts that still intentionally use them

Validation

cd /Volumes/LEXAR/repos/lossless-claw-644-maintenance-remap
./node_modules/.bin/vitest run --dir test \
  test/engine.test.ts \
  test/assembler-blocks.test.ts \
  test/bootstrap-flood-regression.test.ts \
  test/prune.test.ts \
  test/regression-2026-03-17.test.ts

cd /Volumes/LEXAR/repos/lossless-claw-644-maintenance-remap
pnpm run build

Observed locally:

• 5 files passed
• 319 tests passed
• build passed

Stack placement

This is the last downstream migration slice in the current LCM vNext stack.

Related stack:

• umbrella: Martian-Engineering/lossless-claw#641
• frontier base: Martian-Engineering/lossless-claw#646
• continuity base: Martian-Engineering/lossless-claw#642
• fidelity base: Martian-Engineering/lossless-claw#643
• tracked issue: Martian-Engineering/lossless-claw#644
• upstream typed projection seam: openclaw/openclaw#79905

Issue

Refs Martian-Engineering/lossless-claw#644.