fix(beta): capture HTTP status + error into sweep_runs.notes

[Augustas11]

Problem

When a sweep cell fails, sweep_runs records n_err > 0 but the notes column is NULL — the per-request HTTP status and error string already captured by harness.fire_stream are dropped on the floor in aggregate_cell. This caused a real misdiagnosis last week: a ctx=2000 failure through the production gateway was misread as a "gateway streaming read-idle timeout bug" when gateway-path timeouts are all 280–360s. The actual cause was almost certainly a transient 503 provider_unavailable on a marginal single-slot provider, but the harness gave no way to confirm without re-running.

Base is spike/provider-model-autotune because that is where beta/sweep.py lives today; it is not yet on main.

Fix

• In aggregate_cell, collect up to 3 distinct (http_status, error[:80]) pairs while iterating per-request results, join them into a ~200-char summary, and expose it as notes (NULL when every request succeeds). The cell-row builder uses agg["notes"] instead of the hardcoded None. ≤20-line change; no schema migration needed (notes column already exists).

Verification

Local smoke against beta/mock_llm_server.py.

100% error mock (per the verification plan in the task prompt) — notes populated:

context_target  n_err  notes
--------------  -----  ---------------------------------------------
1000            1      HTTP 500: {"error": "simulated server error"}

Green mock (0% errors) — notes stays NULL:

context_target  n_ok  n_err  notes
--------------  ----  -----  -----
1000            1     0

• python3 -m py_compile beta/sweep.py beta/harness.py — OK
• Sweep against erroring mock exits 1; sweep against green mock exits 0.
• No changes outside beta/sweep.py. harness.py, schema, and other phases untouched.

🤖 Generated with Claude Code