SQLite "fetch failed" 500 Error - Minimal Reproduction

When making SQLite calls via https://esm.town/v/std/sqlite after a long-running operation (like an AI call), the request fails with:

Error: {"statusCode":500,"error":"Internal Server Error","message":"fetch failed"}

The SQLite "fetch failed" error occurs when:

1. An HTTP request starts
2. A long-running operation happens (AI call taking 30-120+ seconds)
3. THEN a SQLite call is attempted
4. The SQLite call fails, possibly due to connection/context becoming stale

• Platform: Val Town
• SQLite Library: https://esm.town/v/std/sqlite
• Original val: dcm31/github-leads-wizard

1. Open this val's HTTP endpoint: Click here to open
2. Select a delay time (30 seconds simulates typical AI call time)
3. Click "Run Test"
4. Wait for results
5. If you see ❌ FAILED with "fetch failed" in the error column, the bug is reproduced!

The SQLite INSERT should succeed regardless of how long the request has been running.

After a delay (simulating AI processing), the SQLite call fails with:

{"statusCode":500,"error":"Internal Server Error","message":"fetch failed"}

at createResError (https://esm.town/v/std/sqlite?v=6:81:10)
at Object.execute (https://esm.town/v/std/sqlite?v=6:47:11)
at eventLoopTick (ext:core/01_core.js:179:7)
at async createSession (https://esm.town/v/dcm31/github-leads-wizard@103-main/db.ts:40:3)
at async Array.<anonymous> (https://esm.town/v/dcm31/github-leads-wizard@103-main/main.ts:140:23)

In the github-leads-wizard val:

1. User submits a company to analyze via /api/start
2. generatePrompt() calls AI (takes 30-120+ seconds)
3. createSession() is called which does an SQLite INSERT
4. The INSERT fails with the fetch error
5. This blocks the email results flow

For programmatic testing:

curl -X POST https://dcm31--32c77ddefb9811f0be9942dde27851f2.web.val.run/api/test \
  -H "Content-Type: application/json" \
  -d '{"delay": 30000}'

• This appears to happen consistently after long delays
• May be related to Val Town's request timeout or SQLite connection management
• The original code worked when AI calls were faster