diff --git a/audits/nightly-2026-05-19.md b/audits/nightly-2026-05-19.md
new file mode 100644
index 00000000..691d6c0f
--- /dev/null
+++ b/audits/nightly-2026-05-19.md
@@ -0,0 +1,226 @@
+## Nightly audit — 2026-05-19
+
+Automated audit of the ASA codebase against the project rule that **Phase 1 DSP
+measurements are ground truth** and Phase 2 / Phase 3 must never override or
+write back to Phase 1 values.
+
+Run from worktree `claude/quizzical-jemison-22a55f` against `origin/main`
+(`adb715ef`). Backend reachable on `127.0.0.1:8100` during the smoke run
+(probed via `/openapi.json`); no separate backend was started by the
+Playwright `webServer` block.
+
+### Headline counts
+
+  1. Unit tests: **428 total, 426 pass, 0 fail, 2 skipped** (vitest)
+  2. Playwright: **46 total, 37 pass, 8 unexpected fail, 1 skipped, 0 flaky**
+  3. Phase boundary violations: **0 confirmed**, **0 new review items**
+
+Total issues: **8** (0 unit, 8 e2e timeouts, 0 boundary).
+
+The three decision-gate suite failures noted in the 2026-05-16 audit no longer
+appear — commit [`e9e97e1e`](https://github.com/slittycode/ableton-sonic-analyzer/commit/e9e97e1e)
+("decision-gate skip, gerund exceptions, …") added skip handling so they
+fall under the 2 skipped count instead of failing. The `estimatePlr` boundary
+review item from 2026-05-16 §4 was resolved by commit
+[`adb715ef`](https://github.com/slittycode/ableton-sonic-analyzer/commit/adb715ef)
+("close estimatePlr audit item") with an inline comment documenting why the
+fallback is not a competing source of truth — verified read-only in this
+audit.
+
+---
+
+### 1. Unit tests
+
+All 426 active vitest suites pass. The 2 skipped are the two decision-gate
+multi-model comparators that are correctly gated on the presence of
+`/tmp/decision_gate_*.json` snapshots.
+
+### 2. Playwright failures (8)
+
+All eight failures are timeouts on tests that share the same Vite dev server
+(`webServer: reuseExistingServer: true`, `npm run dev:local`). The pass/fail
+distribution per spec file:
+
+  1. `error-states.spec.ts` — 6 pass / 2 fail
+  2. `file-validation.spec.ts` — 7 pass / 1 fail
+  3. `phase2-degradation.spec.ts` — 4 pass / 0 fail
+  4. `responsive-layout.spec.ts` — 5 pass / 0 fail
+  5. `theme-shell.spec.ts` — 1 pass / 0 fail
+  6. `ui-details.spec.ts` — 10 pass / 2 fail
+  7. `upload-estimate-phase1.spec.ts` — 0 pass / 1 fail
+  8. `upload-phase1-live.spec.ts` — 0 pass / 1 fail
+  9. `upload-phase1-midi.spec.ts` — 3 pass / 1 fail
+  10. `upload-phase1.spec.ts` — 1 pass / 0 fail
+  11. `upload-phase2-live-gemini.spec.ts` — 1 skipped (gated on `RUN_GEMINI_LIVE_SMOKE`)
+
+Failing tests:
+
+  1. [error-states.spec.ts:339](apps/ui/tests/smoke/error-states.spec.ts:339)
+     "error banner dismiss button removes the error" — 90 s timeout. Fully
+     mocked path (`stubEstimateRoute` + `page.route('**/api/analysis-runs')`
+     fulfilled with HTTP 500).
+  2. [error-states.spec.ts:464](apps/ui/tests/smoke/error-states.spec.ts:464)
+     "stop monitoring during interpretation preserves completed measurement
+     without showing an error banner" — 90 s timeout. Fully mocked.
+  3. [file-validation.spec.ts:341](apps/ui/tests/smoke/file-validation.spec.ts:341)
+     "load demo track fetches demo.mp3 into the upload flow" — timeout
+     "while setting up page". Stubs `**/demo.mp3` and backend routes.
+  4. [ui-details.spec.ts:359](apps/ui/tests/smoke/ui-details.spec.ts:359)
+     "top metric cards show TEMPO, KEY SIG, METER, CHARACTER after analysis"
+     — 90 s timeout. Mocked via `stubRoutes(page)`.
+  5. [ui-details.spec.ts:395](apps/ui/tests/smoke/ui-details.spec.ts:395)
+     "header shows SonicAnalyzer brand" — 90 s timeout. The body is just
+     `page.goto('/', { waitUntil: 'networkidle' })` and
+     `expect(page.getByText('SonicAnalyzer')).toBeVisible()`; no upload, no
+     stubs, no backend dependency.
+  6. [upload-estimate-phase1.spec.ts:13](apps/ui/tests/smoke/upload-estimate-phase1.spec.ts:13)
+     "upload shows estimate and local DSP processing copy before phase1
+     completes" — 90 s timeout. **Same test that failed in the 2026-05-16
+     audit; untouched since.**
+  7. [upload-phase1-live.spec.ts:44](apps/ui/tests/smoke/upload-phase1-live.spec.ts:44)
+     "live backend phase1 renders results without connectivity errors" —
+     `page.waitForResponse` exceeded 35 s waiting for
+     `POST /api/analysis-runs`. The skip-gate
+     (`backendSupportsPhase1Routes`) did not skip the test, so a backend
+     was reachable but did not complete the silence.wav analysis in time.
+     This test is the only one in the suite that needs a real backend.
+  8. [upload-phase1-midi.spec.ts:84](apps/ui/tests/smoke/upload-phase1-midi.spec.ts:84)
+     "phase1 dual-source session musician panel renders both blocks
+     simultaneously" — 90 s timeout. Mocked via `stubGeminiPhase2` and
+     explicit `**/api/analysis-runs/*` route stubs.
+
+**Suspected cause:** the seven mocked-path failures share `webServer.reuseExistingServer: true`
++ `await page.goto('/', { waitUntil: 'networkidle' })`. Under sequential load
+against a single shared Vite dev server, the `networkidle` condition is known
+to hang when Vite's HMR client keeps a long-poll connection open or when one
+of the lazy-loaded chunks (e.g. `AnalysisResults.tsx`) is still being built on
+first hit. Test #5 ("header shows brand") is the simplest possible page-load
+assertion in the suite and still times out, which points at server/route
+stability rather than at any production code change. Failures #1–6 and #8 all
+do the same `page.goto('/', { waitUntil: 'networkidle' })` step. The seven
+counterparts in the same files (e.g. the other 10 passing tests in
+`ui-details.spec.ts`) succeed, so this is intermittent.
+
+The eighth failure (#7, `upload-phase1-live.spec.ts`) is independent — it
+needs a backend that completes a real analysis pass on silence.wav within
+35 s, which the host did not meet at run time.
+
+**Suggested next steps (do not auto-fix per audit policy):**
+
+  1. Swap `waitUntil: 'networkidle'` for `'domcontentloaded'` in the
+     smoke specs, or add an explicit selector wait, to remove Vite-HMR
+     long-poll sensitivity. The five `networkidle` waits are at
+     [error-states.spec.ts:340/465](apps/ui/tests/smoke/error-states.spec.ts:340),
+     [file-validation.spec.ts:350](apps/ui/tests/smoke/file-validation.spec.ts:350),
+     [ui-details.spec.ts:360/396](apps/ui/tests/smoke/ui-details.spec.ts:360),
+     [upload-estimate-phase1.spec.ts](apps/ui/tests/smoke/upload-estimate-phase1.spec.ts),
+     [upload-phase1-midi.spec.ts:84](apps/ui/tests/smoke/upload-phase1-midi.spec.ts:84) — confirm
+     against each file.
+  2. Make `upload-phase1-live.spec.ts:44` either (a) gate on a
+     wall-clock budget appropriate for the local hardware, (b) inject a
+     pre-analyzed run id via the backend testing surface and skip the
+     analyze step, or (c) move it out of the default `test:smoke` glob
+     into the same live-only config as `upload-phase2-live-gemini.spec.ts`.
+  3. Re-check [upload-estimate-phase1.spec.ts:13](apps/ui/tests/smoke/upload-estimate-phase1.spec.ts:13)
+     — it has now failed in two consecutive nightly runs (2026-05-16 and
+     2026-05-19) with the same symptom. The 2026-05-16 audit suggested
+     looking at `git blame` on
+     [analysisRunsClient.ts](apps/ui/src/services/analysisRunsClient.ts) /
+     [analyzer.ts](apps/ui/src/services/analyzer.ts) for an interpretation-fetch
+     gating change; the recommendation still stands.
+
+### 3. Phase boundary audit — code paths checked
+
+Phase 2 / Phase 3 code surveyed for mutation, recomputation, or write-back
+into Phase 1 authoritative fields (tonal balance, dynamics envelope, stereo
+image, LUFS / LRA / peak, spectral centroid / rolloff):
+
+  1. [apps/backend/server_phase2.py](apps/backend/server_phase2.py) — read-only.
+     `_normalize_measurement_result_for_gemini` (line 742) builds a shallow
+     `dict(payload)` copy and, for each per-stem entry, a `dict(stem_entry)`
+     copy. Only spectral-field keys are renamed for prompt consistency;
+     values are not changed. `_build_descriptor_hooks` (line 832) reads
+     `measurement_result.get(...)` and emits a separate
+     `MEASUREMENT_DERIVED_DESCRIPTOR_HOOKS` JSON block — does not write
+     back. `_fix_grammar_in_record` (line 1037, modified in
+     [`89df12e1`](https://github.com/slittycode/ableton-sonic-analyzer/commit/89df12e1))
+     mutates Phase 2 recommendation strings only; no Phase 1 path.
+  2. [apps/backend/server_phase1.py](apps/backend/server_phase1.py) — read-only.
+     `_normalize_spectral_detail` (line 81) renames keys on a copy.
+  3. [apps/backend/sample_generation.py](apps/backend/sample_generation.py),
+     [apps/backend/sample_theory.py](apps/backend/sample_theory.py),
+     [apps/backend/sample_synthesis.py](apps/backend/sample_synthesis.py),
+     [apps/backend/sample_drums.py](apps/backend/sample_drums.py),
+     [apps/backend/server_samples.py](apps/backend/server_samples.py) — Phase 3 audition-sample
+     code. Only `.get()` reads from the Phase 1 payload, all citations
+     declared explicitly via `phase1_fields=[...]` lists. No
+     recomputation of LUFS / spectral / dynamics fields.
+  4. [apps/ui/src/services/phase2Validator.ts](apps/ui/src/services/phase2Validator.ts) — read-only.
+     All Phase 1 access via property reads (`phase1.bpm`, `phase1.key`,
+     `phase1.lufsIntegrated`, `phase1.spectralDetail?.spectralCentroidMean`,
+     etc.). The validators check Phase 2 against Phase 1, never the
+     reverse. The recent diff
+     ([`adb715ef..origin/main`](https://github.com/slittycode/ableton-sonic-analyzer))
+     adds an `audience` field on violations and tags the `NEW_FIELD_UNCITED`
+     warning as `dev` — no field-write changes.
+  5. [apps/ui/src/services/mixDoctor.ts](apps/ui/src/services/mixDoctor.ts) — read-only.
+     `estimatePlr` (line 112) and the newly inspected `estimateMonoCompatible`
+     (line 121) both early-return when the explicit Phase 1 field is
+     populated, and only derive locally for the advisory score; results
+     are not written back into the snapshot. `estimateMonoCompatible`
+     reads `phase1.monoCompatible` then falls back to
+     `phase1.stereoDetail.subBassMono` — both are pre-emitted Phase 1
+     fields, not recomputations.
+  6. [apps/ui/src/services/analysisRunsClient.ts](apps/ui/src/services/analysisRunsClient.ts),
+     [apps/ui/src/services/analyzer.ts](apps/ui/src/services/analyzer.ts),
+     [apps/ui/src/services/appliedRecommendations.ts](apps/ui/src/services/appliedRecommendations.ts),
+     [apps/ui/src/services/backendPhase1Client.ts](apps/ui/src/services/backendPhase1Client.ts) —
+     no Phase 1 mutations. `backendPhase1Client.ts` normalizes nullable
+     booleans (`monoCompatible`, `bpmAgreement`, `bpmDoubletime`) on the
+     way *in* from the wire envelope — that is the parsing boundary, not
+     a write-back from Phase 2/3.
+  7. [apps/ui/src/services/sessionMusician/](apps/ui/src/services/sessionMusician/),
+     [apps/ui/src/services/midi/](apps/ui/src/services/midi/) — read-only consumption
+     of Phase 1 / interpretation outputs.
+  8. [apps/ui/src/components/AnalysisResults.tsx](apps/ui/src/components/AnalysisResults.tsx),
+     [apps/ui/src/components/MeasurementDashboard.tsx](apps/ui/src/components/MeasurementDashboard.tsx),
+     [apps/ui/src/components/MixDoctorPanel.tsx](apps/ui/src/components/MixDoctorPanel.tsx) —
+     render-only consumption of LUFS, true peak, stereo stats, spectral
+     balance, centroid, and rolloff. No assignments to those fields anywhere
+     in `src/`.
+
+Strict regex `(phase1|measurementResult|measurement_result)\.[a-zA-Z_]+\s*=[^=]`
+across `apps/ui/src` and `apps/backend/{server_phase2.py,sample_*.py,server_samples.py}`
+returns **zero matches**. The only match for that pattern repo-wide is
+[test_sample_generation.py:106](apps/backend/tests/test_sample_generation.py:106),
+which is a test fixture mutating a local `phase1` dict before passing it to
+the sample-generation harness — not production code.
+
+Phase 3 audition-sample generation (`sample_generation.py` etc.) is shipped;
+the `patchSmith.ts` Phase 3 synth-patch generator referenced in
+[BACKLOG.md](BACKLOG.md) has not landed yet — there is no `src/phase3/`
+directory and nothing in the frontend to audit for that line.
+
+### 4. Needs human review
+
+None. The 2026-05-16 review item (`estimatePlr`) closed on
+[`adb715ef`](https://github.com/slittycode/ableton-sonic-analyzer/commit/adb715ef);
+the comment block at [mixDoctor.ts:102-111](apps/ui/src/services/mixDoctor.ts:102)
+explains the byte-identity argument with the analyze_core.py reference.
+
+---
+
+### What was *not* found
+
+  1. No Phase 1 field assignments anywhere in `apps/ui/src/` or the
+     Phase 2/3 backend modules.
+  2. No mutations of objects returned from the backend `phase1` envelope.
+  3. No backend code path that writes back into `measurement_result` /
+     `phase1` from the Phase 2 worker or Phase 3 sample generator.
+  4. No drift between the citation contract (`phase1Fields`) and the
+     `collectPhase1FieldPaths` walker in `phase2Validator.ts`.
+
+The chain-of-custody invariant is intact at the source level. The eight
+Playwright failures are independent of that invariant — seven are
+`networkidle`-wait flakiness against a shared Vite dev server, one is a
+real-backend timeout on `silence.wav` analysis.