data: stranded vendor evidence prose and bundle templates become data (W49)

[open-coder-ai]

What this changes

No behaviour differs: this externalizes text per the owner standard of 2026-09-02 (org-plan plan/coding-standards.md §2). (1) The per-vendor evidence prose stranded in allow_semantics.py, packaging_limits.py, permissions_render.py and bundler.py moves into schema-validated data tables (data/allow-semantics.json, data/packaging-limits.json, data/permissions-content-reasons.json), each note keeping its stated basis/date as structured fields; the Python modules become loaders with unchanged public names and shapes. (2) The bundle header/runtime/binding source held as string literals moves to data/templates/*.py.tmpl, loaded via importlib.resources and rendered by __TOKEN__ + str.replace, so every template is valid Python exactly as committed and ruff check . lints it in the existing CI lint job (extend-include, documented F821 per-file-ignore for spliced names; formatter excludes templates because their text is frozen into generated bundles). AGENTS.md gains the compact externalized_text hard rule beside code_comments.

Claim check

No adapter or matrix change; every evidence note moved verbatim. Basis/date fields were added only where the prose itself states them (cursor's 2026-08-28 live experiment, gemini_cli's 2026-08-28 source read, the "Documented"/"Source"/"Docs-basis only" phrasings, claude_code's content-rule citation read 2026-08-29) — nothing invented, and rows whose prose states no basis carry none.

• No capability claim is widened without a mechanism behind it
• Any new/changed MATRIX row carries a verified record — n/a, no matrix change
• Payload shapes come from a primary source — n/a, no payload change

Checks

• pytest -q passes — 1446 passed, 4 skipped (33 new tests: tests/test_data_tables.py schema + mutation coverage, tests/test_bundler_templates.py token coverage)
• ruff check . and ruff format --check . pass (the check now also lints the three templates as Python)
• Runtime path is still stdlib-only (importlib.resources only; tests/check_stdlib_only.py green)
• Commits are signed off (git commit -s)

Additional gates for this move:

• Golden fixtures byte-untouched: git diff tests/fixtures is empty.
• bundle() byte-identical for all 12 agents: sha256 of every bundle compared before/after the refactor, equal; the bundle-equivalence and subprocess-replay suites are green.
• Package-data: the derivation test now walks data/ recursively (it previously checked only the top level), so data/templates/*.py.tmpl and data/vendors/*.json are both derived into the wheel manifest check; pyproject.toml ships the new pattern.
• Token coverage is mutation-tested: an orphaned __TOKEN__ planted in a template fails 14 tests; the planted-lint-error check confirmed ruff check . really discovers .py.tmpl files. These mutations fail; no claim the class is closed.
• No version bump (CHANGELOG Unreleased entry only).

Notes for the reviewer

• The two vouch-evidence strings bundle() splices into a generated runtime's degrade() docstring moved verbatim to allow-semantics.json vouch_notes — they are vendor evidence, but they are also frozen bundle bytes, hence verbatim.
• Kept in code as genuinely agentseam-design text (not vendor facts): the condensed allow_semantics.py module docstring stating the bare-ALLOW rule (the audit detail now lives in the data/schema descriptions), the renderer Unrepresentable reason strings in permissions_render.py/permissions.py/packaging.py fallbacks (diagnostics beside the condition that raises them, per the standard's stated line), and tools/recount's stated tables (recount is the derivation source for vendor JSON; externalizing it is a separate decision).
• docs/design/dialect-families.md cites bundler_templates.py line numbers from D3; those references describe the design at its pin and were left untouched.

🤖 Generated with Claude Code

---

Generated by Claude Code