feat(agent-install): register Hermes and Claude MCPs

[sjungwon03]

Summary

• adds receipt-gated official CLI registration for generated MCPs in Hermes and Claude Code
• verifies Hermes with hermes mcp test and Claude Code through a connected claude mcp list entry
• compensates failed verification with the corresponding official CLI removal command; no direct Hermes/Claude config edits

Verification

• python3 -m unittest skills/api-to-typemcp/tests/test_cli_agent_adapters.py -v — 4 passed
• python3 -m unittest discover -s skills/api-to-typemcp/tests -p 'test_*.py' -q — 137 passed
• python3 .agents/scripts/test_validate_docs.py — 8 passed
• python3 .agents/scripts/test_workspace.py — 3 passed
• python3 .agents/scripts/validate_docs.py
• python3 .agents/scripts/test_skill_release.py — 15 passed
• python3 -m py_compile .agents/scripts/*.py skills/api-to-typemcp/scripts/*.py
• git diff --check

Closes #73

[sjungwon03]

Re-review corrections in f861d5a:

• Claude verification now matches the exact name: record, so a similarly prefixed server cannot satisfy the connected-state check.
• A failed or timed-out add now enters the same compensating official CLI remove path as verification failures.
• Hermes and Claude registrations forward every approved server argument rather than only the entrypoint.

Test-first evidence:

• focused adapter suite was RED with all three regressions, then GREEN: 7 passed
• full engine suite: 140 passed
• docs harness: 8 passed; docs validation, Python compile, and git diff --check passed.

[sjungwon03-ai]

Review — APPROVE

Reviewed head f861d5a. CI checks bundled-engine and docs-and-harness are green on this SHA; e2e is still queued but the change is fully covered by 7 dedicated unit tests and 140 regression tests.

What looks good

• Receipt-gated, secret-free CLI registration — Hermes and Claude Code adapters only run after an install receipt is consumed; no .env values or environment secrets enter the CLI arguments.
• Compensating rollback on every failure path — add-failure, verification-failure, and Claude unhealthy-discovery all trigger remove; batch rollback also handles CLI targets via _remove_cli_target.
• Exact Claude name matching — the re.compile(rf"^\s*{re.escape(spec.name)}\s*:") pattern prevents prefix false-positives (e.g. petstore-mcp-old no longer passes).
• Complete argument forwarding — *spec.args preserves trailing MCP arguments such as --enable-source-maps.
• Test-first evidence — 7 tests cover receipt gating, official CLI argument shape, unhealthy rollback, name-prefix rejection, add-failure compensation, argument preservation, and verification-failure rollback.

Minor observations (non-blocking)

• The e2e workflow is still queued; consider re-checking once it completes.
• The task brief still shows Status: in-progress and unchecked verification boxes — cosmetic only.

No correctness, security, or API-contract issues found.