fix(server): Claude adapter v2 fixes (upstream G2)

[rynfar]

Ports upstream T3 group G2 (Claude adapter v2 fixes) into Pylon. Part of upstream cycle #996.

Sources

Upstream SHA	Behavior	Outcome
1e7c8e0f24	Claude MCP tools show their names and icons (tool_use_meta → title + toolSource)	Adopted, clean cherry-pick -x (includes claude_mcp_tool_presentation replay fixture)
b4d3d51ac9	Claude model/setting changes no longer kill running background agents (refuse replacement with ClaudeBackgroundWorkBlocksQueryReplacementError; continuations keep the live process)	Adopted, clean
8283b481a3	Claude V2 turns start on Windows with the default binary path (resolveClaudeSdkExecutablePath)	Adopted, clean (helper already exists in Pylon)
5bf19d12b0	Threads stay working while Claude starts a wake turn (init counts as wake evidence)	Adopted, clean
ce90eec1ff	A Claude command you stop shows as interrupted, not failed	Adopted, clean
aad732901e	Claude threads no longer stuck in plan mode Claude entered itself (setPermissionMode restore on reuse)	Adopted, clean
6414268859	Fake Claude CLI exits with its parent	Already covered: Pylon's fake never had the keep-alive setInterval (Pylon's probe only sends initialize, no get_usage), so it already exits when stdin closes. Intent folded into the adapted fixture below.
858b98b8fa	Fake Claude CLI is a fixture file, not a generated string	Adapted: moved Pylon's own fake (initialize-only) to apps/server/src/provider/Layers/testing/ClaudeCapabilitiesProbe.fixture.mjs; upstream's get_usage handler and keep-alive not carried since Pylon's probe does not request usage.

Pylon adaptations / interactions

• fix(provider): harden OpenCode session titles, v2 serve line, and Claude interrupt grace #916 Claude interrupt grace: that grace (settleInterruptedTurn / CLAUDE_INTERRUPT_GRACE) lived in the v1 apps/server/src/provider/Layers/ClaudeAdapter.ts, which was removed when Pylon adopted orchestrator v2 (eed7e978f6). None of these sources touch v1 or the v2 interruptTurn path; v2 interruptTurn still awaits query.interrupt acknowledgement before close. b4d3d51ac9 only adds a stopping = true mark in closeLiveQueryForNativeThread (Stop/rollback/fork), so Stop still always closes the process. Behavior otherwise unchanged; whether the v1 "wait for the turn to settle before kill" grace should be re-ported to v2 is a separate follow-up (not in scope here).
• The new user-facing error message carries no T3 branding. ProviderFailure.ts maps the new tag to its own message.
• Rebased onto feat: subagent and tool presentation fixes (upstream G3) #999 (G3, merged first). The one conflict was the tool-item title chain: kept G3's dynamicToolTitle(...) and added the MCP presentation.title after it, matching upstream's final order (searchTitle ?? dynamicToolTitle ?? presentation.title ?? null).

Applicability to other adapters (follow-ups, not ported)

• MCP name/icon (1e7c8e0f24): Claude-specific tool_use_meta. No other v2 adapter sets toolSource; Codex/OpenCode MCP presentation is a possible follow-up if their protocols expose server names/icons.
• Model change kills background work (b4d3d51ac9): Claude-specific (selection change replaces the CLI process). Codex/OpenCode take per-turn model selection without process replacement; Prime not checked in depth — follow-up.
• Windows binary path, wake init, self-entered plan mode: Claude CLI/SDK-specific; no equivalent in Codex/OpenCode/Prime.
• Stopped command shows interrupted (ce90eec1ff): Codex, Cursor, Pi and OpenCode2 already assert interrupted in turn_interrupt_mid_tool/opencode2_interrupt. OpenCode v1 (OpenCodeAdapterV2.ts maps tool part error → failed regardless of turn.interrupted) likely has the same bug — follow-up. Prime: not verified — follow-up.

Verification

• vp run -F t3 typecheck — exit 0, 0 error TS (ran tsc --noEmit in apps/server).
• vp test run src/orchestration-v2/Adapters/ClaudeAdapterV2.test.ts src/provider/Layers/ClaudeCapabilitiesProbe.test.ts — 2 files, 131 tests passed.
• vp test run src/orchestration-v2/testkit/ClaudeReplayFixtures.integration.test.ts src/orchestration-v2/testkit/OrchestratorReplayFixtures.contract.test.ts src/orchestration-v2/testkit/OrchestratorReplayFixtures.integration.test.ts — 3 files, 129 tests passed (incl. claude_mcp_tool_presentation, claude_background_task_interrupt, turn_interrupt_mid_tool/claudeAgent).
• Tests run with CLAUDE_CONFIG_DIR unset (known local env issue for claude_result_is_error).
• vp lint / vp fmt --check on changed files — format clean; only two pre-existing no-unused-vars warnings in ClaudeAdapterV2.ts/.testkit.ts.

Re-run after the rebase onto #999: typecheck exit 0; the five test files above together pass (260 tests).

Unverified

• Windows binary path behavior (no Windows host).
• Live Claude CLI behavior for tool_use_meta, plan-mode restore, and background-work refusal beyond the recorded replay fixtures and unit tests.

🤖 Generated with Claude Code

[github-actions]

Thread transfer impact

✅ Thread transfer remains within every enforced ceiling.

ℹ️ The exact PR base did not have a successful artifact. Baseline uses the latest successful main measurement shown below.

Provider	Metric	Main baseline	This PR	Impact	PR ceiling
Codex	Total thread wire	4.9 KiB	4.9 KiB	0 B (0.0%)	6.8 KiB	✅
Codex	Thread snapshot wire	3.7 KiB	3.7 KiB	0 B (0.0%)	4.9 KiB	✅
Codex	Live turn WebSocket wire	1.2 KiB	1.2 KiB	0 B (0.0%)	2.0 KiB	✅
Codex	Live turn WebSocket decoded	20.4 KiB	20.4 KiB	0 B (0.0%)	29.3 KiB	✅
Codex	Live turn messages	2	2	0 (0.0%)	8	✅
Claude	Total thread wire	4.9 KiB	4.9 KiB	0 B (0.0%)	6.8 KiB	✅
Claude	Thread snapshot wire	3.7 KiB	3.7 KiB	0 B (0.0%)	4.9 KiB	✅
Claude	Live turn WebSocket wire	1.2 KiB	1.2 KiB	0 B (0.0%)	2.0 KiB	✅
Claude	Live turn WebSocket decoded	20.8 KiB	20.8 KiB	0 B (0.0%)	29.3 KiB	✅
Claude	Live turn messages	2	2	0 (0.0%)	8	✅

Baseline: cda6633 · PR result: a1175cb · Source CI: success

Scenario and decoded snapshot size

10 historical turns, 5 command tools per turn, 878.9 KiB retained MCP result per historical turn, and a 1.05 MiB retained result in the measured turn.

• Codex decoded thread snapshot: 106.1 KiB
• Claude decoded thread snapshot: 106.4 KiB

Updated in place by a trusted workflow. PR artifacts are strictly validated and never executed.

[rynfar]

CI note: on head a1175cb, Test Server 2 failed once in src/provider/acp/AcpSessionRuntime.processTree.test.ts > terminatePosixOwnedProcessTree > rotates more than 64 live parents without scanning retained tombstones. This PR does not touch that code. It passed when only the failed job was re-run, and the same job passed on the pre-rebase head, so I'm treating it as flaky. All checks are now green.