D4: fold the four F2 flat_decision adapters onto the engine

[open-coder-ai]

What this changes

gemini_cli, tabnine, junie and grok are no longer hand-written adapter modules: the D3 engine executes their data/vendors/<agent>.json entries (dialect-families.md §6, wave D4), and bundle() composes engine + inlined VENDOR literal for them. Wire output through the dispatcher is unchanged — the golden fixture suite passes byte-for-byte with zero fixture edits.

Claim check

• No capability claim is widened without a mechanism behind it — no MATRIX row changes; every new config gate (junie's PermissionRequest) was recounted by replaying the real dispatcher against the frozen D1 payload, and test_config_gates_are_a_subset_of_matrix_block_capable_events still gates it
• Any new/changed MATRIX row carries a verified record — no matrix rows touched
• Payload shapes come from a primary source — no new payload shapes; every renderer string in the new config sections is verbatim from the deleted adapters and frozen in tests/fixtures/golden/<agent>.json

Checks

• pytest -q passes (1348 passed, 4 skipped; includes the golden replay and the 12-agent test_bundler_subprocess.py replay)
• ruff check . and ruff format --check . pass (ruff check --select I --fix was a no-op)
• Runtime path is still stdlib-only (PYTHONNOUSERSITE=1 python tests/check_stdlib_only.py clean; examples/generate.py --check and docs/assets/gen_brand_assets.py --check also run clean)
• Commits are signed off (git commit -s)

Notes for the reviewer

Two-stage history, each commit green. Stage 1 adds the schema fields, engine devices and tools/recount tables and regenerates the four entries via recount_vendor_config.py --write (modules still bound, so the enrichment is derived, not hand-typed — F1 and singleton entries regenerate byte-identical). Stage 2 flips _CONFIG_DRIVEN, deletes the four modules (465 lines), and removes the transitional gemini fields override.

Every engine device is an opt-in config key, absent from all F1 entries (checked: every F1 G1 gate has honours_escalate: false, no F1 entry carries the new keys), so F1 wire cannot move: escalate word path in _g1; hook_specific_tool_input (gemini, emits tool_input: null on a rewrite with no input, as the fixture pins) and top_level_updated_input (junie) transform bodies; template degrade notes filling (reason-or-default, wire event) (gemini's "cannot prompt from a hook"); gate_reason_defaults (junie's Stop "not finished"); allow_context_key/allow_silent_events (junie's additionalContext and silent Stop-allow); missing_wire: reverse_map (tabnine answers hand-built events, and a hand-built UNKNOWN event stays silent); a fields.tool_input envelope chain (grok's toolInput, derived by the recount walker from the deleted module's own ti = assignment); content_only_for_write_tools (gemini); literal entry_extra values (tabnine's name: agentseam).

One designed test retirement. test_the_unverified_vocabulary_is_still_only_tabnine grepped per-vendor source files that no longer exist; dialect-families.md §5's table names it as dying at this wave with its intent re-homed — the re-homed derivation (test_vendor_config.py::test_vocabulary_basis_is_unverified_only_for_tabnine) already exists and passes.

Two respond()-direct behaviours did not survive, both unreachable through handle() (called out in the CHANGELOG): tabnine no longer answers the undocumented BeforeModel/AfterModel wire names (they map to no canonical event, so handle() short-circuited before respond(), and a config gate for them would fail the matrix-subset test), and gemini_cli no longer honours a rewrite at gates the recount classified deny-only (through the dispatcher, degrade() already converted those to escalates). W38's handback note also asked for a flat_entries hook-entry style for tabnine — tabnine's wrapper is actually hooks_map with a name extra, so only the literal-extra device was needed; flat_entries rendering stays for the wave that folds its real users (vscode_copilot/windsurf, D5).

Bundle sizes (design §7 updated with the same numbers): gemini_cli 458→690, tabnine 409→664, junie 429→664, grok 439→664 (+51–62%, above the §4 estimate of 470–500 because an engine bundle inlines the whole shared engine, G2 included — the trade D3 already made; D6 owns bundle composition). Source tree: −465 adapter lines, +~40 engine lines, no new modules.

---

Generated by Claude Code