docs: add Q&A audit log for 2026-03-28 session

signal-loom-ai-AUDIT-LOG-2026-03-28.md

@@ -0,0 +1,101 @@
+# Signal Loom AI — Session Q&A Audit Log
+## Date: 2026-03-28 | Session: Main (Aster Vale / Traves Brady)
+## Session ID: 71a84ded-b60c-4a65-8d03-1e56a81d0483
+
+---
+
+## Q&A LOG
+
+### Q: What is the root cause of "Job still queued" error despite job being completed?
+**A:** `result_json` was being returned by `row_to_job_payload()` but is NOT a field in `Job.__init__`. When `Job(**payload)` was called in `_get_job_or_404()`, Python raised `TypeError: unexpected keyword argument 'result_json'`. This exception was silently caught somewhere, causing the code to fall through to the in-memory `jobs` dict — which still had the job with `status=QUEUED`.
+
+**Files Fixed:**
+- `routers/transcribe.py`: `_get_job_or_404()` now uses `inspect.signature()` to filter payload keys to only those accepted by `Job.__init__`
+- `core/job_store.py`: `row_to_job_payload()` already did not include `result_json` — confirmed clean
+
+**Verified Working:** 3 completed jobs (78f38eb5, 0134f260, c0a061f6) all return correct transcript results.
+
+---
+
+### Q: Why did the worker exit after processing exactly one job?
+**A:** Root cause was `yaml` module was imported inside a function (`process_job`) but `import yaml` at the module level was missing. Worker crashed with `NameError: name 'yaml' is not defined` on the second job's processing. Additionally, the `jobs._persist()` was writing to the `jobs` table (which uses truncated 13-char IDs) instead of `job_registry` (which uses full UUIDs), causing silent failures.
+
+**Files Fixed:**
+- `run_local_worker.py`: Added `import yaml` at module level
+- `run_local_worker.py`: Completion block now writes directly to `job_store.update_job()` (SQLite with full UUIDs) AND to `jobs.update()` (in-memory) as dual-write strategy
+- `jobs/__init__.py`: `_persist()` now writes to `job_registry` instead of `jobs` table
+
+---
+
+### Q: Why is YouTube download extremely slow on the Mac Mini?
+**A:** The network from the Mac Mini has poor/no connectivity to YouTube CDN. yt-dlp hangs on metadata fetching for YouTube URLs. This is a network routing issue, not a code bug. Audio file uploads bypass this bottleneck entirely.
+
+**Status:** Not resolved. Workaround is to use direct audio file uploads instead of YouTube URLs.
+
+---
+
+### Q: Why did LinkedIn video fail on the HF Space?
+**A:** LinkedIn-hosted videos are not directly downloadable — LinkedIn does not expose a direct media URL (unlike YouTube/TikTok). LinkedIn videos require authentication or a session cookie to access. Hugging Face Spaces run on servers that cannot authenticate to LinkedIn.
+
+**Status:** Known limitation. Not a bug in Signal Loom AI. Workaround: download the LinkedIn video locally first, then upload the audio file.
+
+---
+
+### Q: What is the current state of the distributed pipeline?
+**A:**
+- ✅ API server: Running on port 18790
+- ✅ Local worker: Running as background process (PID 47607, since 4:43PM)
+- ✅ Redis queue: Jobs flowing correctly (9 completed, 0 pending)
+- ✅ SQLite persistence: Full UUIDs in `jobs` table
+- ✅ Result API: `/v1/result/{job_id}` returns correct transcripts
+- ✅ YouTube download: Slow (network issue)
+- 🔒 Modal GPU: Needs GitHub OAuth auth (account has $30 credits)
+
+---
+
+### Q: What was discussed about Modal GPU deployment?
+**A:** Traves confirmed the Modal account is set up with $30 in credits. Authentication is required via GitHub OAuth (the Aster-SGLAI GitHub account). Browser relay was used to navigate the OAuth flow, but the GitHub session was not cached in the browser. Traves was asked to run `modal token new` on his Mac Mini terminal to complete the authentication.
+
+**Status:** Pending authentication.
+
+---
+
+### Q: What was the LinkedIn post about?
+**A:** The LinkedIn post (chrischow91) was a Duolingo marketing case study — titled "Let's talk about Duolingo's unhinged marketing #strategy and why it works." It discusses 5 tactics Duolingo uses: (1) Entertain not sell, (2) Localisation done right, (3) Jumping on trends, (4) Building brand personality, (5) Cultural relevance. This was NOT a video about Signal Loom AI — Traves may have been sharing it as a marketing reference/example.
+
+---
+
+## DECISIONS MADE THIS SESSION
+
+1. **Result API bug**: Fixed by filtering `Job.__init__` valid keys in `_get_job_or_404()`
+2. **Worker persistence**: Worker now dual-writes to `job_registry` (SQLite, full UUIDs) AND `JobsManager` (in-memory)
+3. **yaml import**: Added to `run_local_worker.py` module imports
+4. **Modal GPU**: Auth pending — `modal_whisper.py` ready to deploy
+5. **start_worker.sh**: Created shell launcher for worker
+
+## FILES CREATED OR MODIFIED THIS SESSION
+
+- `/signal-loom-api/routers/transcribe.py` — Result endpoint fix
+- `/signal-loom-api/jobs/__init__.py` — _persist() fix
+- `/signal-loom-api/core/job_store.py` — row_to_job_payload() verified clean
+- `/signal-loom-api/run_local_worker.py` — yaml import + dual-write
+- `/signal-loom-api/start_worker.sh` — worker launcher script (NEW)
+- `/signal-loom-api/com.signalloom.worker.local.plist` — launchd plist (NEW, has issues)
+
+## CURRENT STATE
+
+- API: http://localhost:18790 ✅
+- Worker: Background process, waiting for Redis queue jobs ✅
+- Redis: 9 completed, 0 pending, 0 processing ✅
+- Modal: Auth required (Traves running `modal token new`) 🔒
+
+## COMPLETED TRANSCRIPTIONS (VERIFIED)
+
+| Job ID | Source | Segments | Language |
+|--------|--------|----------|----------|
+| 78f38eb5-a5d2-4ef4-8150-8163ea97316b | YouTube (Rick Astley) | 68 | en |
+| 0134f260-cce8-438b-96af-735fd6fed048 | YouTube | 110 | en |
+| c0a061f6-9a65-4e37-b5eb-dae34ac16f1c | YouTube | 124 | en |
+
+---
+*Logged: 2026-03-28 17:04 PDT | Aster Vale for Signal Loom AI*