daily: append late session notes (16:46-17:41 PDT)

memory/2026-03-28.md

@@ -237,3 +237,49 @@ Fix: `row_to_job_payload()` no longer includes `result_json`. Also added `inspec
 - `run_local_worker.py`: Added `yaml` import, dual-write to job_registry + JobsManager
 - `start_worker.sh`: Created shell launcher for worker
 - `com.signalloom.worker.local.plist`: Launchd plist (has issues, shell wrapper used instead)
+
+## 2026-03-28 — Late Session (16:46–17:41 PDT)
+
+### Modal GPU Deployment — Still Blocked on Auth
+- Browser relay attempted: Google OAuth (astervale94@gmail.com) — Google account NOT linked to Modal ("A user wasn't found with this sign-in method")
+- GitHub OAuth also not cached in browser relay session
+- Modal account confirmed funded ($30 credits)
+- FIX REQUIRED: Traves needs to run `modal token new` on Mac Mini terminal to complete GitHub OAuth
+- `modal_whisper.py` ready; deploy command: `modal deploy modal_whisper.py`
+
+### LinkedIn Video on HF Space — Expected Failure
+- HF Space URL: `huggingface.co/spaces/Signalloom-ai/Signalloomai`
+- LinkedIn videos are NOT directly downloadable — LinkedIn requires authentication/session cookies
+- HF Spaces servers cannot authenticate to LinkedIn
+- This is a known limitation, not a bug
+- Workaround: download LinkedIn video locally first, then upload audio file
+
+### Traves Shared: Duolingo Marketing Case Study (LinkedIn)
+- Post by Christopher Chow (chrischow91) re: Duolingo's "unhinged" TikTok marketing strategy
+- 65% DAU increase, 45% revenue rise (late 2023)
+- 5 tactics: entertain not sell, localization, trend-jumping, brand personality, cultural relevance
+- Shared as potential marketing reference for Signal Loom AI — NOT a product issue
+
+### Session Audit Log
+- Saved to: `signal-loom-ai-AUDIT-LOG-2026-03-28.md`
+- Key Q&A decisions documented:
+  1. "Job still queued" bug: `result_json` TypeError in `_get_job_or_404()` — fixed with inspect.signature() key filtering
+  2. Worker exit after one job: missing `yaml` import + `_persist()` writing to wrong table
+  3. LinkedIn video failure: authentication requirement, not a bug
+  4. YouTube slow: network issue, not code bug
+
+### Git Commits (this session)
+- `9090899` — daily: memory + Q&A audit log
+- `437ff7c` — docs: add Q&A audit log for 2026-03-28 session
+- `6395a5c` — reframe: whitepaper abstract updated with new core thesis
+
+### System State at Session End
+- API: ✅ Running (PID 48070, port 18790)
+- Worker: ✅ Running (PID 47607, since 4:43PM, idle waiting for Redis queue)
+- Redis: 0 pending, 0 processing, 9 completed
+- Modal: 🔒 Needs `modal token new` from Traves
+
+### Critical Bugs Fixed Today (Summary)
+1. Result API "still queued" despite completion → `inspect.signature()` filtering in `_get_job_or_404()`
+2. Worker crash after one job → `yaml` import added; `_persist()` fixed to use `job_registry`
+3. `_persist()` was writing to `jobs` table (truncated 13-char IDs) instead of `job_registry` (full UUIDs)