providers: delete dead hal0.voice impl; document lemond-only STT/TTS dispatch

[thinmintdev]

Closes #620

Scope (minimal, per maintainer decision)

Delete only the dead local provider implementation — the in-process
classes that ran Moonshine/Kokoro — while keeping moonshine and
kokoro as valid capability-provider identifiers. lemond (and the
toolbox images) serve STT/TTS now, so the in-process hal0 provider
classes had no live importers and are pure dead code.

Deleted (the #620 win — dead code, no live importers)

• src/hal0/voice/ — the entire re-export shim package (__init__.py, kokoro.py, moonshine.py)
• src/hal0/providers/kokoro.py — KokoroProvider impl
• src/hal0/providers/moonshine.py — MoonshineProvider impl
• tests/providers/test_kokoro.py, tests/providers/test_moonshine.py — tests for the deleted impls

Kept VALID (capability layer untouched — these stay valid provider identifiers)

• config/schema.py _VALID_PROVIDERS — still includes moonshine + kokoro (a SlotConfig(provider="kokoro") validates fine)
• slots/state.py SELF_MANAGED_PROVIDERS — still {kokoro, moonshine, vibevoice}
• capabilities/config.py default provider = "kokoro", capabilities/catalog.py, api/routes/backends.py + models.py classification — all unchanged

Mechanically forced by the deletion

• providers/__init__.py — drop the kokoro/moonshine imports + _PROVIDERS registrations (the runtime classes are gone; the identifiers remain valid config values). get_provider("kokoro") now raises KeyError.
• tests/providers/test_lemonade.py — assert live providers resolve and the retired voice classes raise KeyError from get_provider().

Doc (the #620 doc ask)

• tests/test_import.py — remove the now-deleted hal0.voice module from the import parametrization.
• ARCHITECTURE.md — document lemond-only STT/TTS dispatch, while explicitly noting kokoro/moonshine remain capability-provider identifiers served by lemond/toolbox (not by an in-process hal0 provider class).

Tests / lint

• tests/config/test_schema.py tests/test_import.py → 59 passed (the two failures the broad version introduced — test_all_valid_providers stays green because moonshine/kokoro remain valid; hal0.voice import entry removed)
• tests/slots/ tests/providers/ → 265 passed (no regression from reverting the over-broad enum/SELF_MANAGED changes)
• ruff check → all checks passed; ruff format --check → all formatted

Deferred (not in this PR)

comfyui is not in config/schema.py _VALID_PROVIDERS even though installer/etc-hal0/slots/img.toml declares provider = "comfyui". This is a pre-existing latent gap on main (not introduced or fixed here). Flagging for a separate issue to keep this PR minimal-scope.

🤖 Generated with Claude Code

[thinmintdev]

REVIEWER VERDICT: APPROVE — conditional on green CI (python checks currently pending; the deleted/modified test files are exactly what those checks validate, so certify on green). Adversarial dangling-ref pass is clean.

SPEC (#620): All 5 ACs met.

• Moonshine + Kokoro providers + voice/ shims removed: providers/{moonshine,kokoro}.py deleted; entire voice/ package (__init__.py, kokoro.py, moonshine.py) deleted; dropped from _PROVIDERS, __all__.
• Live exceptions retained: ComfyUIProvider, FLMProvider kept in registry; LlamaServerProvider intact.
• _VALID_PROVIDERS reconciled — comfyui gap fixed (added), moonshine/kokoro removed. (This is the issue-mandated fix, not scope creep.)
• test_legacy_providers_still_registered rewritten → test_live_providers_registered + test_retired_voice_providers_not_registered; test_kokoro/test_moonshine deleted.
• ARCHITECTURE documents lemond-only dispatch + the 3 live exceptions + pi-coder-is-v0.4-gated note.

STANDARDS / adversarial grep: ZERO live importers of providers.moonshine/providers.kokoro/MoonshineProvider/KokoroProvider/hal0.voice anywhere in src/. Remaining moonshine/kokoro string hits are capability keys / lemond recipe names / migration maps / forward-compat enum values — not provider-class usage. Dispatcher has no voice refs. STT/TTS path intact (lemond whispercpp/kokoro recipes).

Non-blocking findings (maintainer option, do NOT gate):

1. Behavioral asymmetry: llama-server/flm stay in _VALID_PROVIDERS and round-trip-as-ignored, but moonshine/kokoro now hard-ValidationError (schema.py:427 provider_valid). #620 explicitly calls the path "vestigial, not a supported runtime" so this is intended — but a user-authored legacy slot TOML with provider="moonshine" will now fail to load. No shipped/seed TOML does this (verified). If smoother legacy upgrades are wanted, keep-as-ignored (like llama-server/flm) is the safe pattern. Flagging for the merge-gate advisor.
2. Stale docstrings still list MoonshineProvider/KokoroProvider as examples: providers/base.py:108, providers/lemonade.py:23.

Merge-ready on green. ARCHITECTURE.md overlaps #649 (see report) — sequential merge needs rebase of the 2nd.

[thinmintdev]

Converted to draft — needs a scoping decision.

Deleting the hal0.voice provider classes is fine (no live importers). But this PR also makes moonshine/kokoro invalid SlotConfig providers, which the capability layer still depends on: src/hal0/capabilities/config.py:17 defaults provider = "kokoro", capabilities/catalog.py maps tts/stt capabilities → moonshine/kokoro, and config.py schema still lists them. Hence test_all_valid_providers → ValidationError (red on previously-green main).

Decision needed: (A) keep moonshine/kokoro as valid capability-provider identifiers (lemond/toolbox serve them) and delete only the dead local impl — minimal, unblocks now; or (B) fully remove them, which requires migrating the capability defaults/catalog off those names. Reviewer read it as intended-(B); the capability-layer refs say (B) is bigger than this PR. Holding for the call.

[thinmintdev]

REVIEWER VERDICT: APPROVE — conditional on green CI (python/γ pending; the deleted-test changes are what those checks validate — certify on green). Re-review of the MINIMAL re-scope (squashed commit 0c1e72c, 11 files). Confirmed truly minimal, NO capability-layer regression. All 6 verification points pass:

1. Dead impls deleted, no live importers — voice/{__init__,kokoro,moonshine}.py + providers/{kokoro,moonshine}.py all gone. Zero live imports of hal0.voice/providers.moonshine/providers.kokoro/MoonshineProvider/KokoroProvider in src/; the only remaining mentions are docstrings/removal-notes (providers/init.py:15,41; base.py:108; lemonade.py:23). ✓

2. _VALID_PROVIDERS UNCHANGED — still {lemonade, llama-server, flm, moonshine, kokoro} (schema.py:89). config/schema.py is byte-identical to main (not in the PR diff). So test_all_valid_providers passes legitimately, not by weakening the assertion. ✓

3. SELF_MANAGED_PROVIDERS restored to {kokoro, moonshine, vibevoice} (state.py:124) — vibevoice is back; the prior over-broad version that dropped it is corrected. ✓

4. test_import.py no longer references hal0.voice — the diff removes the "hal0.voice" entry from the import-smoke list. ✓

5. ARCHITECTURE.md states impls retired, identifiers remain — explicitly: "The dead local MoonshineProvider/KokoroProvider implementation classes (the hal0.voice package …) were deleted in #620 — they had no live importers. moonshine and kokoro remain valid capability-provider identifiers in the config/capability layer (SlotConfig.provider, capabilities/config.py, capabilities/catalog.py …); inference is served by lemond (whispercpp + kokoro recipes)." Correct framing. ✓

6. comfyui-gap is genuinely PRE-EXISTING on main — _VALID_PROVIDERS on main lacks comfyui while installer/etc-hal0/slots/img.toml:18 seeds provider = "comfyui". This latent gap exists identically on main and is untouched here (schema.py byte-identical). Correctly deferred to a separate issue, NOT introduced by this PR. ✓

This addresses my prior review's sole non-blocking concern (the behavioral asymmetry / hard-ValidationError on legacy moonshine/kokoro TOMLs): by leaving _VALID_PROVIDERS intact, those configs still load — no upgrade break. Both prior stale-docstring nits (base.py:108, lemonade.py:23) persist but remain non-blocking.

Scope is exactly: delete dead in-process impls + drop their _PROVIDERS registrations + document lemond-only dispatch. No dispatcher, schema, capabilities, or CLI regression. Merge-ready on green.