docs(c04): honest reconciliation — Q1+Q2 not measured, Q3 affirmed

[epappas]

Summary

• Fills in the C04 pre-registration's Outcome section with honest reconciliation: status INCOMPLETE, Q1+Q2 not measured, Q3 affirmed at 100% kept-rate, Q4 partial. Six of eight predicted outcomes marked NOT EVALUABLE because workload mismatch made the comparison structurally invalid through all eleven attempts.
• Closes T-38, T-39, T-40, T-41 in TODO.md and opens P0 ticket T-42 (C04a attempt 12 — clean relaunch after T-41).
• 437 CPU tests still passing — no code change in this PR, docs and backlog only.

What this PR is and isn't

This is the prediction-vs-reality reconciliation that the pre-registration discipline requires when a campaign doesn't reach its planned verdict. Eleven C04a launches between 2026-05-26 and 2026-05-27 produced ten integration-layer fixes (PRs #49-#58) but never produced a comparable goodput dataset against T-37's auto_tune baseline. The Outcome section now records that honestly, including the 3.05x cost overrun (~$9.77 actual vs $3.20 pre-reg estimate).

This is not a campaign re-launch and not an analysis writeup. The analysis writeup at docs/research/references/12-c04-outcome.md is gated on attempt 12 producing usable data — that is T-42's job.

What was actually achieved (citable)

• T-26c L1 surrogate validated at 100% kept-rate (20/20) at attempt 7 — far above the pre-reg's 30% Outcome A3 threshold. The per-FailureKind sub-classifier is doing its job under real serving. Caveat: under shared-GPU contention the gate's KL ceiling auto-calibrated wider than configured (effective ~16 vs configured 2.0); the tight-gate kept-rate awaits a 2-GPU re-run.
• Ten integration-layer fixes landed, each with regression tests:
  • PR fix(launch): expose --image flag in launch_joint_campaign.sh #49: `--image` flag in `launch_joint_campaign.sh`
  • PR fix(launch): expose --model flag in launch_joint_campaign.sh #50: `--model` flag in `launch_joint_campaign.sh`
  • PR fix(campaign-runner): cap reference vllm serve --max-model-len to 4096 #51: reference replica `--max-model-len 4096` cap
  • PR fix(adapter): cap candidate max_model_len in 1-GPU mode + archive stderr #52: candidate `max_model_len` env-var + per-trial stderr archival
  • PR fix(surface): inject --gpu-memory-utilization when env cap set + catalog omits GMU #53: GMU inject parallel to existing clamp
  • PR fix(driver): lowercase --goodput metric names (ttft/tpot/e2el) for vLLM #54: `--goodput` lowercase metric names (TTFT to ttft, E2E to e2el)
  • PR feat(driver): T-38 — rate-down search mirroring auto_tune.sh #55 (T-38): driver rate-search mirroring `auto_tune.sh` rate-down algorithm
  • PR fix(adapter): T-39 — process-group kill for candidate vllm tree #56 (T-39): `os.killpg` process-group kill with SIGTERM to SIGKILL escalation
  • PR fix(driver): T-40 — emit --percentile-metrics + tighten per-bench timeout #57 (T-40): explicit `--percentile-metrics ttft,tpot,itl,e2el` + 1800 to 600s timeout
  • PR feat(driver): T-41 — thread random_input_len/random_output_len through config #58 (T-41): workload params (`random_input_len`/`random_output_len`) through DriverConfig

What's still open

T-42 (P0, in `TODO.md`): C04a attempt 12 — clean relaunch on 2x A100 spot now that the workload matches T-37 Baseline B. Requires user GPU-spend authorization. Estimated $1.50-2.50.

Test plan

• `uv run pytest -q -m "not gpu and not basilica"` — 437 passed in 8.69s
• Pre-reg Outcome section now has Status, headline numbers, reconciliation table, per-Q analysis, bugs+fixes table, what's still open, cost actually spent, artifacts, next-session restart point, and closing
• TODO.md: T-38/T-39/T-40/T-41 rows added to Closed; C04a attempts 1-11 row added to Closed; T-42 row added to P0
• No code changes; no GPU spend