Add operational-value grader for daily-caveman-optimizer

[Copilot]

daily-caveman-optimizer had no deterministic operational-value grader. This adds one, following .github/skills/operational-value-designer/SKILL.md, measuring whether the workflow's conciseness edits actually stick in the repository rather than counting activity or trusting agent self-assessment.

Design

• Intent: recovered from the adoption commit (9b89368509...) — trim .github/aw/.github/agents instruction files for conciseness without losing accuracy, in daily batches of 5.
• Assignment: the round-robin batch depends on external cache state, so per-run reconstruction can't rely on git history alone. Also found that a separate workflow (pr-description-caveman) rewrites merged PR bodies on every merge, wiping out the usual gh-aw-workflow-id/run-id footer marker (confirmed against real PRs [caveman] Optimize instruction verbosity — network.md (2026-09-01) #57777, [caveman] Optimize instruction verbosity — evals.md (2026-08-28) #56727). Assignment instead searches for the run's PR by title prefix, labels, author, and a bounded time window after run start — classified bounded-ambiguity rather than fully deterministic.
• Metric (pr-reduction-durability, direct, [0,1]): fraction of the PR's own claimed line reduction (from its files API) still present at the evidence cutoff. This avoids inventing an arbitrary target-lines threshold (no such convention exists for markdown, unlike Go source), and directly penalizes reverted or regrown trims:

reductionClaimed = sum(deletions - additions) over eligible PR files
if not merged: value = 0
else: value = clamp((reductionClaimed - growthSinceMerge) / reductionClaimed, 0, 1)

• Maturation: fixed 7 days, derived from the workflow's own 3-day PR auto-expiry plus buffer to observe reverts.
• Baseline: attainment-only — no prior automated/manual instruction-trimming activity existed before adoption.
• Diagnostic: batch-size-ratio (files touched / 5), reported separately from the primary value.

Changes

• .github/graders/daily-caveman-optimizer-operational-value.sh — new evaluator (--definition/--metric/--grade-run)
• .github/graders/daily-caveman-optimizer-operational-value-study.json — study record (designOutcome: accepted) documenting the design decisions and uncertainties above
• .github/workflows/daily-caveman-optimizer.md — wires in graders.operational-value
• .github/workflows/daily-caveman-optimizer.lock.yml — recompiled

Original prompt

[operational-value-study:paper-v1] daily-caveman-optimizer

Use the repository's .github/skills/operational-value-designer/SKILL.md to design
and verify one deterministic operational-value grader for github/gh-aw
workflow daily-caveman-optimizer. This is a fresh research run; do not copy another
workflow's metric or contract.

Requirements:

• Recover adoption-time intent and pre-adoption evidence independently.
• Use exactly one direct primary attainment metric in [0,1]. Keep diagnostics separate.
• Treat unavailable evidence as null, never zero. Make maturation explicit and stable.
• Support historical assignment reconstruction when case and event are null.
• If no direct operational metric is defensible, create only the study record below with designOutcome "blocked" and explain why. Never substitute activity, output volume, traces, or agent judgment.
• Change only .github/workflows/daily-caveman-optimizer.md, its generated lock file,
  .github/graders/daily-caveman-optimizer-operational-value.sh, and .github/graders/daily-caveman-optimizer-operational-value-study.json.
• Do not add generated historical reports to the pull request.

Write .github/graders/daily-caveman-optimizer-operational-value-study.json as JSON with this exact top-level shape:
{
"schemaVersion": 1,
"repository": "github/gh-aw",
"workflowId": "daily-caveman-optimizer",
"recordedAt": "ISO-8601 timestamp",
"recorder": "copilot-coding-agent",
"designOutcome": "accepted|rejected|blocked",
"rejectionReason": null,
"classification": {
"domain": "short research domain",
"subdomain": null,
"taskArchetype": "short label",
"interventionType": "short label",
"outcomeType": "short label",
"evidenceDirectness": "direct|proxy|activity",
"evidenceLocality": "same-repository|cross-repository|external",
"measurementLevel": "nominal|ordinal|interval|ratio",
"metricForm": "binary|proportion|rate|distance-to-target|index",
"assignmentDeterminism": "deterministic|bounded-ambiguity|ambiguous",
"intentExplicitness": "explicit|implicit|mixed",
"markdownSpecificity": "exact|bounded|open-ended",
"outcomeObservability": "immediate|delayed",
"causalClaim": "observational|comparative-not-causal"
},
"decisions": [{
"type": "intent|opportunity|assignment|evidence|maturation|metric|baseline|diagnostic",
"selected": "observable selected design",
"alternatives": ["observable rejected alternative"],
"rationale": "concise evidence-backed rationale, not hidden reasoning",
"evidenceReferences": ["commit, issue, PR, or file reference"],
"changedDuringReview": false
}],
"uncertainties": [{
"type": "measurement|assignment|evidence-availability|temporal|baseline|domain-fit|causal-attribution",
"level": "low|medium|high|unknown",
"description": "specific limitation",
"mitigation": null,
"evidenceReferences": []
}]
}

For blocked/rejected designs, rejectionReason must be a non-empty string. Record
observable contract choices and alternatives only; do not record chain-of-thought.

Validation for accepted designs:

1. .github/skills/operational-value-designer/scripts/verify-operational-value-evaluator.sh .github/graders/daily-caveman-optimizer-operational-value.sh
2. gh aw compile .github/workflows/daily-caveman-optimizer.md

Keep the pull request limited to this workflow. The local study runner will replay
history and archive reports after this task completes.