feat(agent): preview post-tool convergence before boundaries

.omo/evidence/20260727-convergence-v24r15-post-tool-preview/QA.md

@@ -0,0 +1,89 @@
+# V24R15 post-tool convergence preview QA
+
+## What was tested
+
+### Focused cross-layer gate
+
+Command:
+
+```bash
+npm run build
+node --test --test-force-exit --test-timeout 120000 \
+  dist/tests/extension/hooks/model-request-patch.spec.js \
+  dist/tests/agent/progress-lease.spec.js \
+  dist/tests/tool/bash-result-display.spec.js \
+  dist/tests/agent/agent-loop-runtime-controls.spec.js \
+  dist/tests/gateway/map-agent-event-runid.spec.js \
+  dist/tests/observability/recorder.spec.js \
+  dist/tests/observability/local-gateway-progress-handoff.spec.js \
+  dist/tests/products/legal-coverage.spec.js
+```
+
+Artifact: `focused-gate.log`
+
+This drove the bounded parser, HookRuntime event restriction, ToolRuntime
+internal transport, Progress Lease planner, AgentLoop ordering, Gateway event
+projection, O1 recorder, a real local Gateway, and the unchanged Legal Coverage
+validator/transaction surface.
+
+### Full repository regression
+
+Command:
+
+```bash
+npm test
+```
+
+Artifact: `full-suite.log`
+
+### Patch hygiene
+
+Command:
+
+```bash
+git diff --check
+```
+
+Artifact: `diff-check.log`
+
+## What was observed
+
+- Focused gate: 103 tests passed, 0 failed.
+- Full repository suite: 325 tests passed, 0 failed.
+- TypeScript build completed in both gates.
+- The synthetic AgentLoop handoff trajectory reached genuine progress with two
+  `progress_boundary_deferred` events and zero forced full compactions.
+- A stale preview stopped with `boundary_preview_unconfirmed` before a third
+  model request was sent and without running a forced full compaction.
+- Replay, repair-only, exhausted-budget, and multiple-forcing-scope previews
+  did not defer the required boundary.
+- The real local Gateway emitted the ordered O1 sequence
+  `progress-boundary/deferred` then the authoritative
+  `progress-lease/v5` confirmation for both deferred boundaries.
+- The real Gateway trajectory finalized with complete model/tool/turn pairing,
+  zero recorder drops, and no fake API key in the observation log.
+- Legal Coverage produced previews for legal-state writes, but not for reads,
+  unrelated writes, or read-only legal CLI commands.
+- `git diff --check` produced no findings.
+
+## Why this is enough for the code gate
+
+The tests cover both halves of the protocol. Positive tests prove the original
+ordering defect is removed in AgentLoop and in a real Gateway runtime. Negative
+tests prove the preview remains advisory: it cannot carry domain payload,
+consume lease state, bypass the handoff budget, handle multiple forcing scopes,
+or survive a mismatched authoritative report. Full-suite coverage checks the
+new hook effect and agent event against every existing product and runtime test.
+
+This evidence completes only the code gate. It does not claim that the 35-minute
+Case 09 task now completes semantically.
+
+## What was omitted
+
+- No real LLM credentials, auth headers, environment dumps, or private corpus
+  content were copied into evidence.
+- No validator, model, Router, Memory, deadline, corpus, or Progress Lease
+  threshold was changed.
+- No V24R14 campaign artifact was modified or rerun.
+- The V24R15 product campaign remains a separate immutable gate after this
+  commit and PR are available to the isolated candidate instance.