perf(bench): collapse bench-sweep.yml matrix on budget axis

[nikolay-e]

Phase 2 of the work in #39: now that run_final_eval --budgets B1,B2,B3,... reuses compute_scored_state across budgets, the workflow can drop the budget matrix axis and let each cell sweep all budgets in one heavy pass.

Current state

matrix:
  method: [ppr, ego, bm25, aider]
  budget: [-1, 0, 8000, 16000, 32000, 64000, 128000]   # this axis
  depth: [-1, 0, 1, 2, 3, 4]
  test_set: [contextbench_verified, polybench500, swebench_verified]

After excludes ≈ 162 cells × ~500 instances. Each cell pays the heavy phase 7×.

Proposed state

matrix:
  method: [ppr, ego, bm25, aider]
  depth: [-1, 0, 1, 2, 3, 4]
  test_set: [contextbench_verified, polybench500, swebench_verified]

≈ 24 cells. Each cell invokes:

python -m benchmarks.run_final_eval \
    --baseline diffctx \
    --winner /tmp/winner.json \
    --manifests-dir /tmp/manifest_one \
    --budgets -1,0,8000,16000,32000,64000,128000 \
    --workers 10 --out "${CELL_DIR}"

(bm25 / aider: omit -1; aider: omit --budgets and keep one budget per cell.)

Dependent changes

1. benchmarks/aggregate_sweep.py — currently expects cell-<method>-b<budget>-L<depth>-<test_set>/<test_set>.checkpoint.jsonl. Needs to also accept cell-<method>-L<depth>-<test_set>/<test_set>_budget_sweep/[L<depth>/]b<budget>.checkpoint.jsonl and emit one row per (cell, budget) pair.
2. benchmarks/cell_metrics.py — needs to handle multi-budget cell output (one summary per budget rather than one per cell).
3. .github/workflows/bench-sweep.yml — matrix collapse + per-method budget list + artifact naming change (cell-<method>-L<depth>-<test_set>).

Acceptance

• Full sweep wall-time < 90 min on Hetzner CCX63 (vs current ~6 h).
• Byte-identical per-(method, budget, depth, test_set) JSONL rows in grand_summary.json between before/after, on the 5-instance smoke manifest.
• Existing partial-sweep artifacts on bench-results/sweep branch remain readable (_parse_artifact legacy fallback).