Adds the override layer described by ADR-0008 Invariant 4:
``<project>/.memtomem/<type>/<name>/overrides/<vendor>.<ext>``, when
present, byte-replaces that vendor's runtime artifact and skips
auto-conversion. PR-C[2/3] activates this for skills only — the
``OVERRIDE_FORMATS`` matrix ships all 9 ``(asset_type, vendor)`` cells
but the resolver gates non-skills with ``_PR_C_ACTIVE_TYPES``; flip the
gate in a follow-up PR to activate agents/commands.

* `_names.py`: ship `OVERRIDE_FORMATS` (9-cell, ADR-pinned location) and
  `GENERATOR_VENDOR` (`<vendor>_<asset_type>` → vendor) so fan-out and
  PR-D's lint/status share one source of truth.
* `context/override.py`: new module with `resolve(project_root, type,
  name, vendor) -> Path | None`. Reads ONLY from project tree per
  Invariant 1 — the docstring explicitly forbids a wiki argument so
  future-self does not accidentally re-couple sync to the wiki.
* `skills.py`: per-vendor hook after `copy_skill` that replaces the
  vendor's `SKILL.md` only. Auxiliary files (`scripts/`, `references/`)
  stay from canonical (Invariant 4 says ``byte-copy that file`` —
  singular).

Tests in `test_context_override.py`:

* `test_resolve_*` — happy path, no-override, unknown vendor, the
  Invariant 1 no-wiki guarantee, and explicit gate tests for agents
  and commands so the enable-day diff is one-line revertable.
* `test_skills_fanout_applies_claude_override_only` — staged override
  per vendor diverges only that vendor's bytes; others stay canonical.
* `test_skills_fanout_applies_all_three_overrides` — full triplet.
* `test_override_only_touches_skill_md_not_scripts` — auxiliary files
  remain canonical even when an override is present.

The pre-existing byte-identical regression test (added in commit 1/3)
still passes — with no overrides on disk every vendor's SKILL.md still
equals canonical byte-for-byte.

Co-Authored-By: Claude <[email protected]>

Adds the user-facing surface for ADR-0008 PR-C[3/3]: a single command
that seeds ``<wiki>/skills/<name>/overrides/<vendor>.md`` from the
canonical SKILL.md so the user's edit starts from a working baseline
rather than a blank file.

* `wiki/override.py`: `seed_override(store, type, name, vendor,
  *, force)` writes the override file (and a `.bak` sibling on
  `--force`). `render_seed_bytes` returns canonical bytes for skills;
  agents/commands raise `NotImplementedError` matching the resolver
  gate in `context/override.py`.
* `cli/wiki_cmd.py`: `mm wiki skill override <name> --vendor <V>`
  with `--force` (overwrite + .bak) and `--editor` (open `$EDITOR`
  after seeding). Exit-code-bearing classified errors for missing
  wiki / missing skill / collision; no traceback leaks. Stdout prints
  three things — the human "Seeded …" line (substring contract), the
  bare absolute path on its own line for shell capture, and a
  commit hint that names the exact `git add` / `git commit` pair to
  run inside the wiki repo.
* `agent` and `command` subgroups are intentionally NOT stubbed —
  click's "no such command" UX is acceptable v1; PR-D adds them
  alongside `diff`/`lint`/`edit`.

Tests cover happy path, refuse + force + .bak, unknown vendor,
missing skill, missing wiki, editor on/off, and the stdout contract
(substring/presence — no strict line ordering so future UX polish
does not churn the test).

Co-Authored-By: Claude <[email protected]>

* `_names.py`: hoist `Layout = Literal["flat", "dir"]` here so agents.py
  and commands.py share one type definition rather than cross-importing.

* `agents.py` / `commands.py`: introduce `canonical_agent_name(path,
  layout)` / `canonical_command_name(path, layout)` as the single source
  of truth for path → name dispatch. Widen `ExtractResult.imported`
  from `list[Path]` to `list[tuple[Path, Layout]]` so the layout tag
  flows through to consumers; the previously-unused `_layout` from
  `_resolve_*_extract_target` is now used (drops the YAGNI complaint).

* `web/routes/context_agents.py` + `context_commands.py`: drop the
  brittle ``path.name == "agent.md"`` heuristic in import responses.
  List + import handlers both go through the helper now; same dispatch
  shape across the surface, single change point if the layout
  semantics ever shift.

* `wiki/override.py`: pre-flight `render_seed_bytes` (which does the
  source check) BEFORE `target.parent.mkdir(parents=True)` so a
  refused call (missing wiki / missing canonical / collision without
  ``--force``) does not leave a half-built ``overrides/`` directory in
  the wiki tree. Rename the unused `vendor` parameter on
  `render_seed_bytes` to `_vendor` (conventional unused-arg signal),
  drop the `del vendor` line.

* `docs/adr/0008-wiki-layer.md`: update the Status header to reflect
  PR-A + PR-B merged and PR-C in flight; expand the PR-C/D table rows
  to mention the dir-layout BC read, the `_PR_C_ACTIVE_TYPES` gate,
  and `mm context migrate` (PR-D scope). The ADR's Decision and
  Invariants are unchanged.

Tests:

* New `test_canonical_agent_name_dispatch_on_layout` +
  `test_canonical_command_name_dispatch_on_layout` exercise the helper
  for both layouts and the literal ``agent.md`` / ``command.md`` flat
  files (the case the brittle heuristic would misclassify).
* New `test_wiki_skill_override_does_not_create_overrides_dir_on_missing_skill`
  verifies the mkdir-then-refuse fix — refused call leaves the
  ``skills/<name>/`` subtree absent.
* Existing `result.imported` consumers updated to unpack the
  ``(path, layout)`` tuples and call the helper.

Out of scope (carry-forward):
- Two-phase write window in `skills.py` override hook (refactor
  `copy_skill` to take an override-aware source).
- `install_agent` / `install_command` concurrency tests — natural
  fit for the PR-D `_PR_C_ACTIVE_TYPES` flip (more fan-out paths
  exercised concurrently).
- Wiki working-tree dirty UX — natural fit for PR-E (web UI surfaces
  the wiki state already).

Co-Authored-By: Claude <[email protected]>