D5: fold cursor, windsurf, antigravity onto the engine + data/vendors config

[open-coder-ai]

What this changes

The three singleton dialect modules (cursor.py, windsurf.py, antigravity.py) are deleted; each vendor is now a small family module — shape-inferred claims/event naming stays code per design §3.3, plus cursor's G4 and windsurf's G5 renderers — bound to the shared accessor and its data/vendors/<agent>.json entry, which now carries the dialect's words and note strings verbatim (docs/design/dialect-families.md D5). Wire behaviour is unchanged: the D1 golden fixtures pass byte-for-byte, never regenerated.

Two stages, mirroring D3/D4:

1. Stage 1 — schema + engine devices + recount carry the singleton dialect data. All new engine devices are opt-in config keys absent from every F1/F2 entry: a generic dotted-path field accessor (toolCall.args.X, workspacePaths[0], top-level edits[].new_string), stringify, empty_object_events, per-gate words_at overrides, flag_note/flag_note_default templates, hook_entry.also_wires, the cursor/flat_entries hook wrappers with group nesting, and a G1 block reroute for an escalate degraded from a transform where the entry names that degradation. windsurf's pre_mcp_tool_use wiring moves out of entry_extra (which _hook_dict would copy into entries verbatim — W41's landmine) into first-class also_wires; its second blocking pre_tool gate joins _EXTRA_GATE_NAMES, recounted by executing the dispatcher against the frozen pre_tool payload. _hook_json.py splits by activity (_payload.py: claims/parse) to stay in the 300-line budget. The three entries are regenerated by tools/recount_vendor_config.py --write, never by hand; the other nine entries are byte-identical under the same run.
2. Stage 2 — the fold: _cursor.py / _windsurf.py / _antigravity.py family modules, _family dispatch by family, module deletions, and bundle() composing engine + family module + inlined VENDOR literal via one binding template for all eleven config-driven vendors. vscode_copilot stays a dialect module (W38's recorded design feedback; folding it is an owner decision).

The config/code line (§3.1) held with no new named probe: both known hard spots folded inside it — windsurf's two-key pre_tool fan-out is wiring data (also_wires), antigravity's inline event mapping is a wire_events inverse fallback for an entry with no events map, and the sizing clause was not needed.

Test intent survives everywhere; two assertions re-homed with their intent intact:

• test_invocation_events_are_unmapped... asserted on the deleted module's _infer_event; it now asserts the same fact through public parse() → UNKNOWN.
• the git-hook test edited adapters/windsurf.py; it now edits the vendor entry, where that dialect string lives.

Measured bundle line counts (base 38023f4 → this PR; recorded in design doc §7): cursor 524 → 883, windsurf 424 → 814, antigravity 442 → 796 (+69% to +92% — the engine-bundle composition trade D3/D4 already made; D6 owns trimming). Source tree: 472 deleted adapter lines replaced by 219 family-module lines, with every word-shaped fact in config.

Claim check

• No capability claim is widened without a mechanism behind it — MATRIX untouched; windsurf's pre_mcp_tool_use gate maps to pre_tool, already matrix block-capable, and is derived by executing the dispatcher, not asserted
• Any new/changed MATRIX row carries a verified record — no matrix rows changed
• Payload shapes come from a primary source — no new payload shapes; every wire fact is replayed from the D1 golden fixtures (tests/fixtures/golden/*.json), which were captured from the pre-fold adapters and are unchanged in this PR

Checks

• pytest -q passes (1348 passed, 4 skipped; golden fixtures byte-for-byte, 12-agent bundler subprocess replay green)
• ruff check . and ruff format --check . pass
• Runtime path is still stdlib-only (tests/check_stdlib_only.py OK)
• Commits are signed off (git commit -s)

Also run locally: python examples/generate.py --check (up to date), cd docs/assets && python gen_brand_assets.py --check (current), tools/validate_vendor_config.py (all entries ok).

Notes for the reviewer

• Mutation checks on the new devices (worker-protocol discipline, run locally, not committed): flipping antigravity's words_at.Stop.allow to "halt", windsurf's escalate note, and cursor's honours_transform each fails the golden-fixture replay; a benign key-reorder of a whole entry still passes. These mutations fail correctly — necessary, not sufficient.
• Two accepted parse-level nuances, invisible on the wire and to every test: the engine's chain accessor returns "" where antigravity's old raw.get("error") or None returned None for a present-but-empty error; and cursor's content chain falls through past an empty edits join where the old _content_of stopped. Flagging in case either matters to a reviewer more than it did to the suite.
• _g1's escalate reroute is data-triggered (escalate_from_transform present + gate honours escalate); no existing F1/F2 entry has that combination, which the byte-identical fixtures confirm.

---

Generated by Claude Code