feat(eval): complete the fresh 24-run trajectory set across all cells

[hsskey]

Why

The fresh 24-run trajectory comparison (prompt 3d7bfc957c1a, model claude-opus-4-8) was incomplete: prior paid attempts each stopped partway through on the same intermittent provider error, leaving unknown-component, variant-only, and recovery without a comparable row in every cell. After this change the set is complete, so the per-condition comparison can be read for every task x repeat.

What changed

Appended one clean attempt's 24 trajectory rows to runs/2026-09-18.jsonl and regenerated reports/trajectory.md from the accumulated records. The supersede-by-attemptId mechanism (merged in #15) replaced the earlier errored and incomparable slots, giving Runs 8 / Incomparable 0 / Missing 0 for each of cli-canonical, cli-agent, mcp-agent. No source changed.

Success per task in this partition (prompt 3d7bfc957c1a / claude-opus-4-8):

Task	cli-canonical	cli-agent	mcp-agent
known-component	2/2 (~28k in)	2/2 (~28k)	2/2 (~28k)
unknown-component	1/2 (~102k)	2/2 (~43k)	2/2 (~43k)
variant-only	2/2	2/2	1/2
recovery	2/2 (~79k)	1/2	2/2 (~66k)

On the hard tasks (unknown-component, recovery) the agent views match or beat the canonical baseline's success while roughly halving input tokens.

How I checked

pnpm tokenloom eval trajectory --matrix eval/trajectory.yaml --budget-usd 5
# 24/24 rows appended, $4.1103, exit 3 (three runs unsuccessful: MAX_TURNS / NO_SUCCESSFUL_TOOL_CALL / NO_TOOL_REQUEST; not a set stop)

pnpm verify
# PASS types patterns tests reference determinism properties mutations rules benchmarks scope scoring encoding selftest (890s)

The regenerated report satisfies the trajectoryEvidence gate: report measurements match the survivor records and no unsupported adoption claim is present.

What I left alone

No gate weakened, no vitest -u, no committed evidence rewritten. runs/*.jsonl is append-only and its committed prefix is byte-identical to the base. No source or adapter change. The verify run's incidental bench/results.jsonl append and timestamped proof record were reverted / left uncommitted, matching prior PRs.

This spends money: the attempt made real Anthropic calls through the claude CLI under --budget-usd 5, cost $4.1103, each invocation under the $5 ceiling. Cumulative for this converge effort is about $5.94 ($1.83 prior + $4.11 here).

Diagnostic note (no fix applied here, for maintainer awareness): the recurring stop that ended the prior attempts is a single claude -p subprocess exiting 1 with only an init line and no result/usage, which the harness maps to MISSING_AUTHORITATIVE_COST and treats as a set-level stop per SPEC 9.6. The captured init line shows the subprocess loading the machine's user-global Claude Code plugins, hooks, and MCP servers; claude --restricted removes only project and local settings, not user-global configuration (its help notes --strict-mcp-config is needed to skip MCP). So each eval call boots a heavy external stack that intermittently fails, and one such failure halts the whole 24-run. This attempt completed 24 calls cleanly, but the fragility remains and may be worth isolating the eval subprocess configuration or not halting the entire set on a single missing-cost call.