feat(eval): claude provider source for the recommendation harness (zero-Gemini-cost comparison)

BACKLOG.md

@@ -4,7 +4,7 @@ Source: `active/sonic-architect-app` (keep in `active/` as reference). Most of t
 
 ## Shipped — Data
 - ✅ `data/genreProfiles.ts` — 35 EDM genre spectral targets, LUFS/crest factor/PLR ranges. **Landed:** [`apps/ui/src/data/genreProfiles.ts`](apps/ui/src/data/genreProfiles.ts).
-- ✅ `data/abletonDevices.ts` — 14 spectral-band → Ableton Live 12 device mappings + FX rule engine. **Landed:** [`apps/ui/src/data/abletonDevices.ts`](apps/ui/src/data/abletonDevices.ts).
+- ✅ `data/abletonDevices.ts` — 14 spectral-band → Ableton Live 12 device mappings + FX rule engine. **Landed:** [`apps/ui/src/data/abletonDevices.ts`](apps/ui/src/data/abletonDevices.ts). *Ported but demoted to a research-only scoring baseline (owner decision 2026-06-11): imported by no product code, consumed only by the recommendation-eval bridge. See the NEEDS-WIRING decision in [`apps/backend/NEEDS.md`](apps/backend/NEEDS.md).*
 
 ## Shipped — Mix Analysis
 - ✅ `services/mixDoctor.ts` + `components/MixDoctorPanel.tsx` — spectral-balance scoring against genre profiles. **Landed:** [`apps/ui/src/services/mixDoctor.ts`](apps/ui/src/services/mixDoctor.ts), [`apps/ui/src/components/MixDoctorPanel.tsx`](apps/ui/src/components/MixDoctorPanel.tsx).

CLAUDE.md

@@ -378,7 +378,7 @@ ASA_MOSS_MODEL_ID=""                     # optional MOSS model id forwarded to t
 ASA_MOSS_SIDECAR_MODE="mock"             # sidecar-side (its own venv): `mock` (default, deterministic schema-valid Phase-1-grounded output) or `model` (a 501 licence-gated stub that executes no OpenMOSS code).
 ASA_CLAUDE_CLI="claude"                  # path to the Claude Code CLI binary used when ASA_PHASE2_PROVIDER=claude; defaults to `claude` on PATH.
 ASA_CLAUDE_MODEL=""                      # optional model override for the claude provider (e.g. `sonnet`); empty = the CLI's default model. The subprocess runs sandboxed: --safe-mode, --tools "", --no-session-persistence, schema enforced via --json-schema.
-ASA_CLAUDE_TIMEOUT_SECONDS="600"         # per-call timeout for the claude provider subprocess (a full producer_summary prompt on the CLI default model measures ~6 min).
+ASA_CLAUDE_TIMEOUT_SECONDS="600"         # per-call timeout for the claude provider subprocess (a full producer_summary prompt on the CLI default model measures ~6 min). For headless runs also set MAX_THINKING_TOKENS=0 — a thinking-enabled model can spend the entire budget deliberating before the structured output starts (see docs/PHASE2_PROVIDER.md, 2026-06-11 addendum).
 ```
 
 Phase 2 is gated by `VITE_ENABLE_PHASE2_GEMINI`. `GEMINI_API_KEY` is backend-only. `SONIC_ANALYZER_ADMIN_KEY` is backend-only and never exposed to clients.
@@ -440,7 +440,7 @@ A quick map from intent to the right place to start:
 
 ## Backport Candidates
 
-The full original `sonic-architect-app` backlog is **shipped** — genre profiles, Ableton device mappings, mix doctor, eight detectors, Phase 3 audition-sample generation, and `patchSmith.ts` (Phase 3 Vital preset generation, [`apps/ui/src/services/patchSmith.ts`](apps/ui/src/services/patchSmith.ts)). See [`BACKLOG.md`](BACKLOG.md) for the landed inventory. Consult it before re-implementing genre detection, mix analysis, acid/reverb/vocal/supersaw/bass/kick detection — they're already in [`apps/backend/analyze_detection.py`](apps/backend/analyze_detection.py) and emit fields visible in `EXPECTED_TOP_LEVEL_KEYS`.
+The full original `sonic-architect-app` backlog is **shipped** — genre profiles, Ableton device mappings (`abletonDevices.ts` — ported but **research-only** since the 2026-06-11 demotion decision: a scoring baseline for the recommendation-proof harness, imported by no product code; see the NEEDS-WIRING decision in `apps/backend/NEEDS.md`), mix doctor, eight detectors, Phase 3 audition-sample generation, and `patchSmith.ts` (Phase 3 Vital preset generation, [`apps/ui/src/services/patchSmith.ts`](apps/ui/src/services/patchSmith.ts)). See [`BACKLOG.md`](BACKLOG.md) for the landed inventory. Consult it before re-implementing genre detection, mix analysis, acid/reverb/vocal/supersaw/bass/kick detection — they're already in [`apps/backend/analyze_detection.py`](apps/backend/analyze_detection.py) and emit fields visible in `EXPECTED_TOP_LEVEL_KEYS`.
 
 ## Companion Agent Docs
 

apps/backend/NEEDS.md

@@ -85,8 +85,9 @@ citations**, so the chain-of-custody penalty drives its adjusted aggregate to
    #2). The rules **are** feature-triggered, so attaching the triggering measurement
    as a citation is a concrete, harness-rewarded improvement — a candidate
    sub-goal-3 "score-driven change" (frontend edit to `abletonDevices.ts` →
-   `npm run verify`). But see the dead-code finding under NEEDS-WIRING before
-   investing in this path.
+   `npm run verify`). Moot since 2026-06-11: the NEEDS-WIRING decision below
+   demoted `abletonDevices.ts` to a research-only baseline, so this candidate is
+   no longer a *product* improvement path.
 
 This is the first real signal toward the Gemini verdict — likely Gemini earns its
 place on citation + full-surface coverage. Confirm on real renders; the synthetic
@@ -169,22 +170,23 @@ factor, onset density) is the app's own projection (`analyzer.ts`); reuse it rat
 than re-deriving, and only then is the deterministic score a real verdict input
 rather than the synthetic illustration above.
 
-**FINDING — the deterministic source is currently dead code (affects sub-goal 3
-framing).** `apps/ui/src/data/abletonDevices.ts` is imported **nowhere** in the UI
-(no component, no test) — verified by grep. ASA's product recommendations come
-entirely from Gemini (`Phase2Result`); the deterministic engine is unwired,
-ported-but-unused. Consequence: a "score-driven improvement" to `abletonDevices.ts`
-(e.g. the citation-emit fix flagged above) would **not reach any user** — it would
-raise a harness number for code nobody runs, which is harness-gaming, not a product
-improvement (PURPOSE.md decision #5). So before sub-goal 3 treats the deterministic
-path as a real recommendation source, decide one of:
-1. **Wire `abletonDevices.ts` into the product** (so its recs — and any citation
-   improvement — actually reach users), then score + improve it; or
-2. **Treat it as a research-only baseline** in the three-source comparison (score
-   it, but land score-driven *product* improvements on the **Gemini** path
-   instead — which needs the `chore/...` branch + renders).
-This is the honest reason no score-driven improvement was "landed" this session:
-the only render-free candidate would have improved dead code.
+**FINDING — the deterministic source is dead code on the product path. ✅ DECIDED
+2026-06-11: demoted to research-only baseline (option 2).**
+`apps/ui/src/data/abletonDevices.ts` is imported **nowhere** in the UI (no
+component, no test) — verified by grep. ASA's product recommendations come
+entirely from the Phase 2 providers (`Phase2Result`); the deterministic engine is
+unwired, ported-but-unused. Consequence: a "score-driven improvement" to
+`abletonDevices.ts` (e.g. the citation-emit fix flagged above) would **not reach
+any user** — it would raise a harness number for code nobody runs, which is
+harness-gaming, not a product improvement (PURPOSE.md decision #5). The owner
+resolved the wire-or-demote choice: **demote**. Standing consequences:
+1. The module stays where it is as the scored free baseline in the three-source
+   comparison (its file header now records the status); it is **not** to be
+   wired into the product.
+2. Score-driven *product* improvements land on the **Phase 2 provider** path
+   instead. The citation-emit candidate above is off the table as a product
+   change (it remains a legitimate baseline-fairness tweak if the comparison
+   ever needs it).
 
 ---
 
@@ -201,8 +203,19 @@ known-settings axes (role recall / value accuracy) — those need rendered fixtu
 so the *complete* "does Gemini recover the actual settings?" verdict remains
 render-gated.
 
-Mechanism is in place: `--source baseline|gemini|deterministic` on the same
-corpus. To deliver the **full** verdict:
+**Claude provider scored on the proxy corpus — 2026-06-11, zero Gemini cost.**
+The `--source claude` harness path plus `scripts/gen_claude_phase2.py` generated
+and scored fully-cited, zero-warning recommendations on the three
+proxy-fingerprint fixtures: **acid 0.485, house 0.424, melodic_techno 0.424**
+vs recorded Gemini 0.172 / 0.343 / 0.000 on identical fingerprints (text-only
+vs audio+JSON — full table and caveats in `RECOMMENDATION_VERDICT.md`). The
+melodic_techno "Gemini 0 recs" outlier is resolved as Gemini-side: Claude
+produced 13 cards from the same fingerprint. Evidence committed as
+`phase2.claude.json` in each fixture dir. Real-render re-runs remain the
+authoritative step for both providers.
+
+Mechanism is in place: `--source baseline|gemini|claude|deterministic` on the
+same corpus. To deliver the **full** verdict:
 1. Render the fixtures (NEEDS-FIXTURE #1–2).
 2. Wire the deterministic adapter (NEEDS-WIRING).
 3. For each fixture, produce a live Gemini `phase2.json` (needs `GEMINI_API_KEY`;

apps/backend/RECOMMENDATION_VERDICT.md

@@ -80,12 +80,65 @@ recs cited and path-valid** against the real fingerprint (custody 1.000), full-s
 coverage. Consistent with the corpus verdict — Gemini's custody advantage is real,
 not a synthesis artifact.
 
+## Claude provider head-to-head (2026-06-11 — same proxy corpus, zero Gemini cost)
+
+The selectable Claude provider (`ASA_PHASE2_PROVIDER=claude`,
+`phase2_provider.ClaudeCliProvider`) was scored on the three fixtures whose proxy
+fingerprints survived the 2026-06-10 corpus re-authoring, via the exact server
+path (`server._run_interpretation_request`) and the same scorer. Generation:
+`scripts/gen_claude_phase2.py`; scoring: `--source claude` (reads the committed
+`phase2.claude.json` evidence in each fixture dir). Model `sonnet`
+(claude-sonnet-4-6), **text-only** (grounds purely on the embedded Phase 1 JSON,
+no audio), `MAX_THINKING_TOKENS=0`, 300–365 s per fixture.
+
+| Fixture | Claude | Gemini (recorded above) |
+|---|---|---|
+| acid_303_128 | **0.485** | 0.172 |
+| house_sidechain_pluck_124 | **0.424** | 0.343 |
+| melodic_techno_arp_124 | **0.424** | 0.000 (zero cards) |
+| **Mean (shared subset)** | **0.444** | 0.172 |
+
+All three Claude results are fully cited (custody penalty 1.000) with **zero**
+`validationWarnings` — both the citation-path check and the Live 12 catalogue
+gates came back clean.
+
+Findings:
+1. **The melodic_techno outlier was Gemini-side, not fixture-side.** From the
+   identical fingerprint where Gemini returned 0 structured cards, Claude
+   produced 13 rec cards + 14 chain cards + 5 workflow steps (32 envelope
+   entries) and scored 0.424. The fixture's fingerprint is interpretable.
+2. **Kick role recall 1.00 on all three** (the recorded gemini/deterministic
+   table above shows kick 0.00/0.00): Claude's kick cards name `Operator`,
+   matching the specs' source synth. Whether genuine inference or
+   electronic-genre convention, it softens the "source instrument is
+   unrecoverable from audio" ceiling — notably, Claude never heard any audio.
+3. The groove/fx/stereo zeros persist for Claude too (cards exist — `Drum Buss`
+   groove cards, a `Hybrid Reverb` that lands in `unknown` domain — but miss
+   the spec devices), consistent with the domain-attribution limitation in
+   `NEEDS.md`. Provider-agnostic and scorer-consistent.
+
+Caveats, honestly: the SYNTHETIC-PROXY warning at the top applies in full. The
+comparison is like-for-like on *inputs* (identical fingerprints, same scorer)
+but not on *modality*: Gemini received the unrepresentative proxy audio
+alongside the Phase 1 JSON, while Claude's text-only grounding may have
+shielded it from proxy-audio artifacts. Model classes also differ (sonnet vs
+gemini-2.5-flash). Re-run both on real renders before treating this as a
+provider verdict. What it does establish: the zero-Gemini-cost route produces
+fully-cited, catalogue-clean, schema-valid recommendations at
+competitive-or-better measured quality on every fixture it has seen.
+
 ## Where real input is needed (retroactive)
 1. **Real Ableton renders** of the 5 specs (48 kHz/24-bit) replace the proxies →
    makes the known-settings axes (role recall, value accuracy) authoritative. The
    proxies have artifacts: dnb BPM read half-time (174→116), `acidDetail` fires on
    all (synth is saw-heavy), melodic_techno key mis-detected.
 2. **`melodic_techno_arp_124` Gemini 0 recs** — investigate the parse/shape outlier
-   on a real render before trusting that fixture's contribution.
+   on a real render before trusting that fixture's contribution. *(Partially
+   resolved 2026-06-11: Claude produced a full 13-card result from the same
+   fingerprint — see the head-to-head section above — so the outlier is
+   Gemini-side, not fixture-side. Still verify Gemini on a real render.)*
 3. The deterministic `AudioFeatures` projection is approximate (not the app's
-   `analyzer.ts`); reconcile when wiring the deterministic path for real.
+   `analyzer.ts`). The wire-or-demote question was resolved 2026-06-11: the
+   deterministic engine is a research-only baseline and will not be wired into
+   the product (see `NEEDS.md`), so reconciling the projection matters only for
+   baseline fairness in this comparison, not for any product path.

apps/backend/scripts/emit_deterministic_recs.ts

@@ -1,8 +1,10 @@
 /**
  * Deterministic recommendation source bridge (GOAL.md sub-goal 3 inner loop).
  *
- * EVAL / RESEARCH ONLY. Wraps the product's deterministic recommendation engine
- * (`apps/ui/src/data/abletonDevices.ts`) and emits the scorer's normalized rec
+ * EVAL / RESEARCH ONLY. Wraps the deterministic recommendation engine
+ * (`apps/ui/src/data/abletonDevices.ts` — itself research-only since the
+ * 2026-06-11 demotion decision; see apps/backend/NEEDS.md) and emits the
+ * scorer's normalized rec
  * shape so `scripts/evaluate_recommendations.py --source deterministic` can grade
  * the *free* path without a Gemini call. Importing the real module keeps a single
  * source of truth — no Python re-port to drift from the TS.

apps/backend/scripts/evaluate_recommendations.py

@@ -9,7 +9,7 @@
 catalog-valid (the always-runnable half of ingest), pulls recommendations from a
 chosen **source**, and scores how well they recover the known device settings.
 
-The ``--source`` switch is the mechanism sub-goal 3 needs to compare three
+The ``--source`` switch is the mechanism sub-goal 3 needs to compare
 recommendation sources on the same corpus:
 
   * ``baseline``      — trivial no-op source (empty rec set). Runs now, no deps.
@@ -20,6 +20,12 @@
                         ``GET /api/analysis-runs/{run_id}/export/phase2`` works.
                         Producing it live needs GEMINI_API_KEY + rendered audio
                         (needs-fixture).
+  * ``claude``        — score a stored ``Phase2Result`` JSON from the Claude CLI
+                        provider (``--phase2`` or a sibling ``phase2.claude.json``
+                        in the fixture dir). Produce it at zero Gemini cost with
+                        ``scripts/gen_claude_phase2.py`` (text-only, grounds on the
+                        stored Phase 1 fingerprint; needs a logged-in Claude Code
+                        CLI, no API key).
   * ``deterministic`` — score the ``abletonDevices.ts`` path. Wiring the node
                         bridge that emits normalized recs from a real Phase 1
                         fingerprint is a documented follow-on (NEEDS.md); until
@@ -82,13 +88,17 @@ def _resolve_recs(
     if source == "baseline":
         return rev.normalize_baseline(fixture), "trivial baseline (empty)"
 
-    if source == "gemini":
+    if source in ("gemini", "claude"):
+        sibling_name = "phase2.json" if source == "gemini" else "phase2.claude.json"
         candidate = phase2_path
         if candidate is None and fixture.source_path is not None:
-            sibling = fixture.source_path.parent / "phase2.json"
+            sibling = fixture.source_path.parent / sibling_name
             candidate = sibling if sibling.exists() else None
         if candidate is None or not candidate.exists():
-            return None, "no Phase2Result JSON (pass --phase2 or drop phase2.json in the fixture dir)"
+            return None, (
+                f"no Phase2Result JSON (pass --phase2 or drop {sibling_name} "
+                "in the fixture dir)"
+            )
         phase2 = rev.coerce_phase2_payload(
             json.loads(candidate.read_text(encoding="utf-8"))
         )
@@ -162,11 +172,12 @@ def run_self_test() -> int:
 def main() -> int:
     parser = argparse.ArgumentParser(description=__doc__,
                                      formatter_class=argparse.RawDescriptionHelpFormatter)
-    parser.add_argument("--source", choices=["baseline", "gemini", "deterministic"],
+    parser.add_argument("--source", choices=["baseline", "gemini", "claude", "deterministic"],
                         default="baseline", help="recommendation source to score")
     parser.add_argument("--fixture", help="limit to one fixture slug")
     parser.add_argument("--corpus-dir", type=Path, default=DEFAULT_CORPUS_DIR)
-    parser.add_argument("--phase2", type=Path, help="Phase2Result JSON for --source gemini")
+    parser.add_argument("--phase2", type=Path,
+                        help="Phase2Result JSON for --source gemini/claude")
     parser.add_argument("--recommendations", type=Path,
                         help="pre-normalized rec list JSON (overrides --source resolution)")
     parser.add_argument("--report", type=Path, help="write a markdown report to this path")