fix: recognize failed V2 ACP tool results as failures

[ranxianglei]

Fixes ranxianglei#426

Problem

The pinned @opencode/plugin@2.0.3 Promise adapter has no safe native error channel, so createV2Tool returns resolved results for caught execution failures and the host records them as state.status === "completed" even when the content says ACP compress failed: .... ACP's internal V2 projection preserved that completed status, so:

• messageHasCompress (lib/messages/query.ts) treated a failed compression as success;
• nudge logic (lib/messages/inject/inject.ts) advanced lastPerMessageNudgeTokens / set compressBaselineSet after a failed compression;
• cold-start reconstruction (lib/state/rebuild.ts) replayed the failed call and rebuilt a block that never existed;
• hideFailedCompressCalls never hid the failed call (it only matches status === "error").

The installed E2E fake-provider classifier also missed these failure texts (ACP .* execution failed does not match ACP compress failed:).

Fix

• errorResult() in lib/v2/tools.ts adds explicit acpFailed: true metadata to every resolved error result (additive — existing per-reason flags unchanged).
• New pure module lib/v2/projection/acp-failure.ts: canonical ACP tool-name list, metadata key, position-0-anchored failure-text pattern, predicates — shared by the projection and the E2E fake provider so they cannot drift.
• toolState() projects a host-completed result for an ACP tool to an internal status: "error" when explicit failure metadata is present or the output matches the historical failure-text pattern (pre-fix host history). It exposes no output field, so provider-owned lowered results are never rewritten: origin correlation keeps the lowered output untouched and the patcher's status-change guard sees identical states on both sides.
• inspectToolResults in scripts/e2e/fake-llm-server.ts classifies ACP results through the shared anchored pattern (legacy regex retained for non-ACP tools).
• installed-v2.ts nudge-growth stage now asserts the observed compress results are genuinely completed, not merely present.

No nudge code changes were needed: messageHasCompressAttempt is status-agnostic (failed attempts still reset pending-nudge state), while messageHasCompress requires completed — once the projection emits error for failures, success baselines stop advancing and V2 aligns with V1 semantics exactly.

Verification

• 18 new tests in tests/v2-failed-tool-recognition.test.ts (pure module, projection shapes, cold rebuild, warm multi-turn nudge baselines with preserveRecentMessages: 20); verified they fail with the projection fix reverted (bug-detection check per AGENTS.md §5.7.3).
• Full suite: 1428/1429 pass. The single failure is tests/soft-block.test.ts, which hardcodes mkdirSync('/tmp/...') at import time and crashes on this sandbox's read-only /tmp mount — environmental, unrelated to this diff (passes in CI).
• Typecheck, build, and Prettier checks pass on the changed files.
• V1 behavior unchanged: recognition is confined to the V2 projection boundary; successful ACP outputs, non-ACP tools, native host error states, and persisted state formats are all untouched.