fix(mcp): let a Task Capsule record its acting Host

[Endofthestars]

Continues the REPAIR work from #21. That round left three items open; this one closes the only one that was actionable here — the Claude Code CLI Tier 1 Host validation — by actually running the trial rather than inspecting the contract.

What the trial found

Read-only discovery passed immediately: a fresh Claude Code CLI 2.1.220 conversation connected to the Benchwork MCP server and discovered all 38 registry tools, with benchwork_status and benchwork_next_actions returning ok: true / mcp-tool-result/1.0 — both from an explicit request and from the broad "What should we do next for this study?" with 26 non-MCP tools still available.

The bounded Task loop did not pass. benchwork_open_task minted every Task Capsule with a hardcoded codex Host and exposed no way for the acting Host to declare itself, so truthful claude-code provenance was refused:

VALIDATION_REJECTED: Agent Result provenance Host does not match its Task Capsule

Athanor was right to refuse. The defect was in the MCP boundary — in a module whose own docstring claims to be host-neutral. Notably the trial Host reported the rejection instead of resubmitting a false codex provenance to force a green run.

This is exactly what the evidence policy means by "contract inspection is insufficient": hosts.py already exported a ClaudeCodeHostAdapter, the Capsule schema already enumerated claude-code, and Ward already admitted it. Only a live Host reached the one line that did not.

The fix

benchwork_open_task now accepts the same optional host_session argument that benchwork_complete_task and benchwork_fail_task already took, and mints the Capsule for that Host.

• Default remains codex, so existing clients behave exactly as before.
• An unregistered Host is refused at the boundary instead of being minted and later rejected by Ward.
• Additive optional field under the MCP API Policy; the 38 frozen tool names are unchanged.

Result

The repaired loop completed end to end, verified independently from Chronicle rather than from the agent's self-report:

Item	Value
Task	TK-B280B5160682, capability bench.code.inspect
Capsule Host	claude-code
Ward	PASS
Completion	ok: true, Receipt RC-47D74468807A
Chronicle	2 verified events, Receipt chain intact

The recorded provenance carries the real --session-id, so it cross-checks against the transcript that produced it. The Host used only Read alongside the Benchwork tools, matching the Capsule Circle. The source repository never acquired a .benchwork/ directory.

Claude Code CLI moves from PENDING_HOST_VALIDATION to PASS. HOST-IDE-001 and WAITING_FOR_DISCLOSURE_AUTHORIZATION are untouched — neither was exercised here.

Gate

178 unittest cases OK, 175 pytest passed + 46 subtests, ruff clean, mypy clean on 35 files, 92 schemas validated, doc links verified, release policy permitted.

Coverage note: main currently measures 79.93% locally, just under the configured fail_under = 80. This branch raises it to 80.01%.

Evidence: examples/phase2-final/host-claude-code/