D6: bundle composition — legacy splice path deleted, engine trimmed per entry

[open-coder-ai]

What this changes

bundle() now has exactly two compositions and no filesystem-probed legacy splice path: bundle_entry(cfg) (public) renders contract + one family-engine section trimmed to what the entry uses + the inlined VENDOR literal + the runtime for all eleven config-driven vendors, and vscode_copilot — the deliberate dialect-module exception — keeps a dedicated composition whose output is byte-identical to the legacy splice it replaces. Docs' "adding an agent" paths now describe the config entry; design §7 records the measured numbers.

Trimming (dialect-families.md §7's open item, owned by D6) is composition logic only — no engine module changed: the family's bound entry points (hj_* / <family>_* + hook_entry_config) plus their transitive dependencies are extracted from the shared modules, minus

• the grammar renderers no gate of the entry speaks (_g2 leaves devin and every F2 bundle; _g1 leaves kimi_code),
• the marker-claims and hj_respond machinery the singleton dialects never call,
• the TOML emitter unless config_format is toml,
• the probe/windows helpers the entry does not cite (unchanged conditions, now inside the one engine section).

Measured sizes (lines, same base — before is main 6838d69, where every engine bundle inlined the whole engine):

bundle	before	after		bundle	before	after
claude_code	786	698		tabnine	760	652
codex_cli	772	685		junie	760	652
kimi_code	760	654		grok	760	652
devin	786	669		cursor	883	608
gemini_cli	786	669		windsurf	814	581
antigravity	796	643		vscode_copilot	503	503 (identical)

The singletons give back most of D5's growth (cursor −31%, windsurf −29%, antigravity −19%). F1/F2 members give back 11–15% but stay above their own wave-base numbers (claude_code 698 vs 650 at D3's base): the D4/D5 devices live inside shared function bodies (_g1, hj_parse), where line-level trimming would mean per-vendor engine variants — §3.1's config/code line says no, so those lines stay, recorded in design §7.

Also: the dead adapters/_common.py (last user deleted in D3–D5) is removed; CHANGELOG Unreleased entry added; no version bump.

New tests (design §5, D6's rows):

• test_bundle_equivalence.py — every bundle replays the whole golden set (every frozen payload × outcome through the bundle's own main(), plus both hook_config paths) — the §5.5 equivalence proof, on top of the existing 12-agent × outcome subprocess replay, which passes identically before and after the trim.
• test_bundler.py — exactly one family-engine section and exactly one VENDOR literal per config-driven bundle (the re-homed intent of the retired cross-adapter splice test, per §5); trimmed-device presence derived from the entry's own gates/config_format; probes inlined exactly where cited.
• test_thirteenth_vendor.py — §5.4: a synthetic vendor added by config alone is schema-validated, bound, claims/parses/responds/wires, and bundles via bundle_entry with the bundle matching the library's wire truth — no code.

Claim check

• No capability claim is widened without a mechanism behind it (no matrix, config, or engine change; the synthetic thirteenth vendor lives only inside its test)
• No new/changed MATRIX row (the test's synthetic row is injected and removed by a fixture)
• No new payload shapes — every replayed payload is a frozen D1 golden fixture

Checks

• pytest -q passes — 1399 passed, 4 skipped (was 1348: +51 new tests); golden fixtures byte-for-byte untouched
• ruff check . and ruff format --check . pass
• Runtime path is still stdlib-only (tests/check_stdlib_only.py green; bundler change is compose-time only)
• Commits are signed off (git commit -s)

Notes for the reviewer

• A trimmed bundle can retain a dead reference: F2's hj_respond still reads if gate["grammar"] == "G2": return _g2(...) with _g2 absent. That branch is unreachable for the inlined entry's data — the same pattern _hook_entry.py's own comment already documents for powershell_command — and the golden-set equivalence test plus the subprocess replay pin the behavior. If you'd rather fail-closed on it, the alternative is rewriting the dispatch in the emitted source, which I judged worse than a provably-dead name.
• Grammar-renderer trimming keys on the entry's own gate table, not a hardcoded family→grammar map, so it is self-maintaining (devin, F1 but G1-only, drops _g2); the engine modules stay fully shared.
• render_config (kimi's TOML emitter) is kept in bundles whose config_format is toml and trimmed elsewhere — bundles never call it at runtime, but it is part of that vendor's config surface.

---

Generated by Claude Code