Summary

PR-D C1a in the ADR-0008 wiki layer series. Removes the PR-C gate
(_PR_C_ACTIVE_TYPES) and lifts wiki/override.py:render_seed_bytes
so agents and commands seed via the vendor generator path that PR-D-prep
(#627) already wired into the fan-out. Skills behavior is unchanged.

PR-D split per the canonical plan (pr-d-glistening-sketch.md):

• C1a (this PR): gate flip + render lift + 4 test inverts + 4 new render_seed_bytes tests.
• C1b (next): mm wiki agent override / mm wiki command override CLI mirror + ~20 mirror tests. Deferred because including it would push C1 past the ~280 LOC target.
• C2/C3/C4: mm context update / status / install --all / migrate (subsequent PRs).

What changed

• context/override.py — drop _PR_C_ACTIVE_TYPES + the gate branch. resolve() now returns paths for every row in OVERRIDE_FORMATS.
• context/_names.py — drop the gate-mention comment; tighten ("commands", "codex") placeholder rationale to point at the runtime NotImplementedError since the gate is gone.
• wiki/override.py — render_seed_bytes lifts the skills-only NotImplementedError. Path (b) parse + render: reuses PR-C's layout-aware parse_canonical_agent / parse_canonical_command + the existing AGENT_GENERATORS / COMMAND_GENERATORS registry rather than widening the generator API surface with a new render_from_canonical(Path) helper. Function-body imports dodge a wiki ↔ context cycle (context.install already imports wiki.store).
• ("commands", "codex") keeps an explicit NotImplementedError since GENERATOR_VENDOR has no codex_commands entry — preserves the diagnostic UX vs a silent KeyError.
• test_context_override.py — invert the 4 gate-pin tests:
  • test_resolve_skips_agents_in_pr_c → test_resolve_returns_path_for_agents_override
  • test_resolve_skips_commands_in_pr_c → test_resolve_returns_path_for_commands_override
  • test_agents_fanout_does_not_apply_override_under_gate → test_agents_fanout_applies_claude_override (3-assertion marker pattern)
  • test_commands_fanout_does_not_apply_override_under_gate → test_commands_fanout_applies_claude_override (same)
  • Each rename keeps the original test name in the docstring for series-PR archive value (PR-C feat(context): PR-C — wiki override + agent/command install widening (ADR-0008) #624 + PR-D-prep feat(context): fan-out override application sites (PR-C carry-forward, ADR-0008 PR-D-prep) #627).
• test_wiki_override.py (new, 4 tests) — render_seed_bytes for agents/codex (TOML), commands/gemini (TOML), the codex commands NotImplementedError placeholder pin, and the traversal-name rejection guard (added in fixup; see below).

3-assertion marker pattern (Step 2 of TDD-ish ordering)

PR-D-prep #627 used a 2-assertion symmetric pin (positive marker present + negative !=). PR-D extends to 3 assertions:

1. canonical body absent (b"Body of the agent." not in body)
2. override marker present (b"OVERRIDE_MARKER_AGENT" in body)
3. byte-equality (body == override_body)

Catches three failure modes: (1) stale canonical leak, (2) wrong override file picked, (3) byte-level corruption. Extension of feedback_pin_invert_symmetric_assertion.

Known limitations (deferred to C1b)

• _dropped fields from vendor renderers are silently discarded.
  Gemini renderer drops tools / skills / model; codex agents drops
  tools / skills / isolation / kind / temperature. Users editing
  the override file won't realize which fields the vendor runtime can't
  represent until C1b surfaces them via stderr WARNING. Inline comment
  in render_seed_bytes flags this as the follow-up.

API change

• render_seed_bytes raises NotImplementedError with a new message
  format for ("commands", "codex") (was: skills-only NotImplementedError
  pre-PR-D, with message "override seeding for {asset_type!r} lands in a follow-up PR"). Internal API — no external callers expected.
• render_seed_bytes parameter rename _vendor → vendor (was unused
  for skills-only path; now load-bearing for vendor dispatch). Internal API.
• render_seed_bytes now calls validate_name on entry (defense-in-depth;
  see fixup below). Idempotent with seed_override's validate.

Behavior change

• Manually placed <project>/.memtomem/<type>/<name>/overrides/<vendor>.<ext>
  files (which had no effect under the PR-C gate) now activate. No CLI
  created these for agents/commands before C1b, so unlikely in practice;
  worth flagging since the gate flip is the activation point.

Fixup commit

fixup(c1a): defense-in-depth validate_name in render_seed_bytes + follow-up notes
(0aca0d4). Self-review caught: render_seed_bytes is in __all__ and
constructs store.root / asset_type / name / ... paths. A traversal-shaped
name from a direct caller (bypassing seed_override) would escape the
wiki root. PR-D C1a is the first PR where the agents/commands branch is
live, so this is the right ship to defend at the function boundary.

• validate_name(name, kind=...) at function entry. Idempotent with
  seed_override's existing validate.
• New test_render_seed_bytes_rejects_traversal_name covers all 3 asset
  types with traversal-shaped names.
• Inline comment about _dropped follow-up.

Test plan

• uv run pytest packages/memtomem/tests/test_context_override.py -v — 13 pass (4 inverted + 9 unchanged).
• uv run pytest packages/memtomem/tests/test_wiki_override.py -v — 4 pass (new file; 3 + traversal guard).
• uv run pytest packages/memtomem/tests/test_wiki_cmd_override.py -v — 10 pass (skills CLI surface unchanged).
• uv run pytest packages/memtomem/tests/ -m "not ollama" -q — 3453 pass, 46 deselected. No regressions.
• uv run ruff check packages/memtomem/src packages/memtomem/tests — clean.
• uv run ruff format --check packages/memtomem/src packages/memtomem/tests — clean.
• Cycle smoke implicit: test_render_seed_bytes_* exercise function-body imports at call time; ImportError would surface there.
• CI green (waiting).

Out of scope

• C1b (CLI mirror) — separate PR, depends on C1a merge.
• C2 installed_at capture move (PR-B contract micro-fixup) — first sub-commit of the C2 PR.
• ADR-0008 status update — final PR in the series.

🤖 Generated with Claude Code