Kill the sticky salvage hard-block's false positives

[TheGreatAxios]

Closes CL-6964. Closes CL-6704. Closes CL-6710.

Three related fixes to the sub-agent stop/salvage machinery in src/subagent/stop-policy.ts and src/subagent/brief-dispatch.ts.

CL-6964 — delete the unreachable thrash salvage machinery

PR #558 removed the re-read thrash stop, so no report can ever carry a "progressive thrash" Summary anymore. Deleted the now-dead isThrashSubAgentReport, appendThrashParentHint/THRASH_PARENT_HINT, the classifyBriefSalvage thrash branch, and the "thrash" entries in HARD_BLOCK_SALVAGES/BriefSalvageKind. Every other salvage kind (no-progress, no-ship, repetition, never-acted, never-edited, turn-budget, deadline, stalled, cancelled, incomplete-report) is untouched.

CL-6704 — stop hard-blocking on Summary substrings

The classifiers matched free-text Summary substrings like "no progress" or "cancelled", so a SUCCESSFUL report whose Summary happened to contain those words was recorded as a forced stop — hard-blocking identical re-dispatch and arming primary salvage nudges for work that actually succeeded.

Fixed by making forcedStopReport's summary strings a single FORCED_STOP_SUMMARIES lookup, shared with the classifiers via isForcedStopSubAgentReport, which now compares with exact equality instead of .includes/.toLowerCase().includes. A typed marker would need a report schema change (that's CL-6786, human-gated, left alone) — exact string matching is the fix that fits here.

False-positive paths now gone: a successful report whose Summary happens to mention "no progress," "cancelled," "long silence," etc. no longer classifies as any forced-stop salvage.

CL-6710 — clear the sticky block after a parallel sibling succeeds

Two concurrent identical-brief task calls can both admit. If one salvaged and the other succeeded, recordOutcome kept the hard-block lastSalvage sticky through the success, so the failing sibling's block later refused a brief that had already produced a good report in the same wave.

Fixed: a successful complete now always clears lastSalvage — the success itself proves the brief is re-dispatchable.

Regressions added

• src/subagent/index.test.ts: successful Summaries containing forced-stop phrases classify as null (no salvage); true forced-stop strings still classify correctly; parallel admit → salvage + success → next admit allowed.
• Existing thrash-labeled ledger tests renamed to use no-progress (the class they were actually generically exercising) since "thrash" is gone from BriefSalvageKind.

Gate

bun run check (lint, typecheck, build, full test suite) green: 5332 pass, 0 fail.

[linear-code]

CL-6964

CL-6704

CL-6710