Record the Cursor event gap (R3), and wire harness + conformance into the probe

[jothimani-rajendran]

What this changes

Two commits, both closing gaps rather than adding surface.

1. R3's note now exists in the repo (closes R3: Cursor needs a one-to-many REVERSE_EVENT_MAP once an ask-style policy ships #73) — on hj_reverse(), the function that maps a canonical event to one wire name, because that map is the limit.
2. harness.py and conformance.py have consumers. Both shipped in Seam conformance: the harness registry, the drift definition, and a derived tier table #146 with nothing calling them. --driver harness builds a probe driver from the registry; agentseam probe conformance feeds recorded measurements to the classifier.

The practical payoff: docs/witness-skeleton.json's three commands no longer contain a hand-typed driver string. All three now read --driver harness.

Claim check

• No capability claim is widened without a mechanism behind it — no MATRIX cell touched, no basis changed, no new recording
• Any new/changed MATRIX row carries a verified record — none added or changed
• Payload shapes come from a primary source — the two Cursor payloads in test_cursor_ask_dialect.py are the shapes already filed in this repo's vendor entry; the ask/deny behaviours were confirmed by executing the adapter, not read off a doc

Checks

• pytest -q passes — 1680 passed, 4 skipped (1656 on main; +24)
• ruff check . and ruff format --check . pass
• Runtime path is still stdlib-only — harness gained os, shlex, subprocess, all stdlib
• Commits are signed off

Notes for the reviewer

R3: checking it sharpened the claim. The plan said "note only, gap already resolved", which reads as closed — but the note was in neither _cursor.py nor docs/. Executing the adapter showed the runtime is already honest: ask at preToolUse returns {"permission": "deny"} with "preToolUse cannot prompt for confirmation, so this is a block", while the same decision at beforeShellExecution returns {"permission": "ask"}. So the gap sits one step earlier than the plan frames it — agentseam install cursor --events pre_tool writes preToolUse and nothing else, foreclosing ask at install time, before dispatch is reached. No code changed (deny-style policies are all that ship, per R3's own terms); six tests pin the facts so the note can't rot into folklore.

A defect in my own last-round code, found by wiring it up. conformance.classify reported a single vendor as AGREED — agreeing with itself, claiming a cross-vendor check that never ran. Same family as the incapable-unanimity bug caught in #146, one row further out. The count that matters is vendors asked, not vendors capable, so one capable vendor beside an excused one is still a real VENDOR_LIMIT; a test holds that distinction so the fix can't over-correct.

Windows quoting. shlex.quote is POSIX-only and these runs happen on a Windows machine as often as not, where a POSIX-quoted argument isn't slightly wrong — it's a different command. list2cmdline on nt, shlex.quote elsewhere.

Opt-in, and the default order is asserted unchanged. --driver harness is expanded at resolution into a real command, so check_record_args, the rendered table and the report's driver field all correctly see a live driver. Whether a run counts as evidence must not shift because the registry gained a row, and a test pins the existing order (recorded if one covers this gate, else reference).

Conformance reports undecidable everywhere today, deliberately. One agent is recorded, so there is nothing to compare, and that looks like an empty feature. Two tests prove the mechanism actually decides once a second recording lands: agreement when a twin agrees, a seam gap and exit 1 when it disagrees. Nothing is fabricated to fill the table meanwhile.

Verified past the suite, since none of this is exercised by parser dispatch. A stub claude on PATH confirms the registry's argv reaches a real process with the prompt intact as a single argument, and the run reports hook_reached: False — the correct reading for a stub that never ran the hook. probe conformance, probe list, probe run --driver harness and tier-table were each dispatched through the CLI.

Unchanged and still yours: the three gaps themselves. They now need only --driver harness --agent-version <version>, and all three are --agent claude_code.

---

Generated by Claude Code