test(core): cover manual compaction cancellation - #45335

Why

Manual compaction must settle a delivered control even if interruption arrives before summary execution begins. It must also leave the inbox available for mutations while the model is running. Existing streaming-cancellation tests did not pin down these two cases.

What Changes

Add two runner regression tests using the existing fixture and TestLLM:

The tests use explicit gates and the test clock rather than wall-clock sleeps. The inbox test times out only its completion observer and always releases the summary gate.

Scope

The earlier production extraction has been removed. This is now test-only: one file, 81 added lines, no new helpers. Runner/Compaction ownership, interruption handling, interfaces, and existing fixtures are unchanged.

Verification

# packages/core
bun run test test/session-compaction.test.ts test/config/compaction.test.ts test/session-runner.test.ts
bun run test test/session-runner.test.ts --test-name-pattern 'after delivery before summary entry|allows inbox cancellation' --rerun-each 10 --concurrent --max-concurrency 4
bun run test
bun typecheck

# worktree root
bunx prettier --check packages/core/test/session-runner.test.ts
bunx oxlint packages/core/test/session-runner.test.ts
bunx ast-grep scan -c script/ast-grep/effect-simplifications/sgconfig.yml --off=unused-suppression packages/core/test/session-runner.test.ts
git diff 2ce126e2df...HEAD --check

• Focused compaction/config/runner suites: 185 passed, 674 assertions.
• New regressions, ten concurrent repetitions each: 20 passed, 100 assertions, against unchanged production code.
• Full Core: 3,768 passed, 39 skipped, 0 failed, 78,802 assertions across 220 files.
• Core typecheck, formatting, Effect scan, and whitespace checks passed. Oxlint reported no errors and three existing warnings.
• An AST audit confirms all pre-existing runner tests and fixture code are unchanged; only the two named regressions were added.
• Cleanup reviews found no actionable reuse, quality, or lifecycle issues. Normal pre-push verification passed all 33 workspace typecheck tasks (27 cached), without bypassing hooks.

Local verification ran on macOS. Fresh CI is running for tests-only head c476515eb1; the previous extraction revision's green CI is not claimed as verification of this head. Auto-merge remains disabled.