Extend Sapat Providers in Daytona

[jjoanna2-debug]

Summary

• Harden the Sapat provider guide around the highest-risk review surface: endpoint control, credential handling, generated artifact cleanup, and bounded async polling.
• Pin the guide to the live Sapat source shape verified against nibzard/sapat main at ba900b7, including provider registry discovery, @register, ProviderConfig, TranscriptionResult, and AsyncPollProvider.
• Add maintainer-ready proof commands and review checklists so a provider adapter can be checked without reconstructing the workspace or exposing private audio/transcripts.

Decisive review points

• Endpoint exfiltration path closed: the guide now requires production traffic on api.assemblyai.com, rejects URL credentials/query strings/unknown hosts, and only allows localhost through an explicit test-only flag.
• Async provider contract pinned: async_poll.py is documented as shared helper-only; concrete providers belong in their own modules and must register provider classes with @register.
• State leakage blocked in tests: registry-mutating tests are called out to reset _registry and _discovered so provider availability does not leak between cases.
• Webhook sprawl rejected: the provider pattern stays on bounded polling unless Sapat grows a signed webhook receiver; no public tunnels, callback URLs, or webhook secrets are introduced just for an adapter.
• Review proof de-noised: validation commands distinguish placeholders from real credentials and check for committed .env, media, and generated transcript artifacts.

Security / privacy hardening

• API keys stay in environment variables or workspace secrets; no keys belong in README examples, screenshots, logs, PR bodies, fixtures, or endpoint URLs.
• Endpoint overrides must not redirect credentials or uploaded audio to arbitrary hosts. Local endpoints are test-only.
• Live validation uses short non-sensitive clips only; private calls, unreleased demos, medical files, legal recordings, generated transcripts, and payout details stay out of commits.
• Provider errors must not log authorization headers, raw .env values, full provider responses, provider request metadata, or full transcript text.
• Temporary MP3 output and smoke-test .txt files are explicitly excluded from committed review artifacts.

Current validation

• npx --yes markdownlint-cli@0.45.0 guides/20260511_extend_sapat_providers_in_daytona.md definitions/20260511_definition_transcription_provider_adapter.md authors/jean_claude_joanna.md
• git diff --check
• file guides/assets/20260511_extend_sapat_providers_in_daytona_img1.png and sips -g pixelWidth -g pixelHeight guides/assets/20260511_extend_sapat_providers_in_daytona_img1.png verified a valid 1600x900 PNG.
• High-confidence token/private-key scan across changed text files returned no hits.
• git ls-files | rg '(^|/)\.env$|\.mp3$|\.mp4$|sample\.txt$' returned no committed smoke-test artifacts.
• git grep -nE 'Bearer [A-Za-z0-9_.-]{20,}|ASSEMBLYAI_API_KEY=' -- . | rg -v 'ASSEMBLYAI_API_KEY=your_|git grep -nE' returned no real credentials.
• Codex Security diff preflight returned status: ready; final diff audit found no credential, private-media, generated-transcript, or executable-runtime expansion in this content-only patch.
• Latest pushed commit: f747241031672c935c644ad9c7fc83d886b93e73; DCO is green.

/claim #13