@productbrain/cli 0.1.0-beta.30 → 0.1.0-beta.3015

package/dist/scoreboard/diagnose.js

@@ -0,0 +1,293 @@
+/**
+ * SYNCED from packages/scoreboard-core/src/index.ts — DO NOT EDIT HERE.
+ * Canonical SSOT for the flywheel scoreboard diagnosis (consumed by BOTH the published
+ * `pb` CLI and the Cortex page). The CLI cannot import the workspace package at publish time
+ * (tsc, no bundler), so it carries this committed copy. Edit the source, then run:
+ *   node scripts/sync-scoreboard-core.mjs
+ * A content-hash drift gate (scripts/check-mcp-lib-drift.mjs) fails CI if this drifts.
+ */
+// User-facing copy carries NO raw Chain IDs (chain-id-leak guard): the CLI prints `cause` /
+// `nextAction` / `reading` verbatim to any workspace's terminal, so an ID here would leak our
+// internal governance refs cross-tenant. The governing entry (the no-flip-threshold decision) is
+// cited in code comments only, never in these strings.
+const UNPINNED = 'threshold unpinned — pin against the live distribution';
+// WP-527 (Lane T Slice 2): PROPOSED (not pinned) floor values, drafted from the observed
+// distributions in docs/evals/PRESS-RELEASE-EVAL.md's Product-OS calibration runs — a proposal for
+// founder accept, never a committed threshold, and never evidence about any OTHER workspace calling
+// this shared function (no workspaceId here — PB is multi-tenant); the `cause` strings say
+// "Product-OS calibration" explicitly so the number is never misread as the caller's own history.
+// Status stays `awaiting` regardless (never ok/degraded from a proposal), per DEC-1282.
+// VISIBILITY GAP: reaches the `pb scoreboard` CLI (formatters/scoreboard.ts renders cause/
+// nextAction) but NOT Cortex — +page.svelte:171 renders only tile.reading + a static tooltip.
+// Surfacing PROPOSED state there is real, disclosed follow-up work, out of WP-527's scope.
+// M3 in-session-ratify: observed 0% in every Product-OS run so far — a conservative early bar.
+const M3_PROPOSED_IN_SESSION_FLOOR = 0.1;
+// M5 has TWO INCOMPARABLE age instruments (diagnoseM5's local-pull drift vs diagnoseM5FromServer's
+// server-generation age) — each gets its OWN single-observation calibration below, never a
+// cross-instrument "trend" (the bug this fixed: Run 1's local-age and Run 3's materialization-age
+// are different measurements, not two points on one series). Local-pull: only on-record reading is
+// 14.1d (Run 1). Server-generation: only on-record reading is ~5.1d (Run 3's materialization age).
+const M5_PROPOSED_LOCAL_AGE_FLOOR_DAYS = 10;
+const M5_PROPOSED_SERVER_AGE_FLOOR_DAYS = 7;
+function pct(n) {
+    return n === null ? 'n/a' : `${(n * 100).toFixed(0)}%`;
+}
+function ms(n) {
+    if (n === null)
+        return 'n/a';
+    const h = n / (1000 * 60 * 60);
+    if (h < 1)
+        return `${(n / (1000 * 60)).toFixed(0)}m`;
+    if (h < 48)
+        return `${h.toFixed(1)}h`;
+    return `${(h / 24).toFixed(1)}d`;
+}
+/**
+ * M1 — retrieval truth-state mix. The no-domain-share flip (INS-1775 floor) is UNPINNED, so the
+ * status is `awaiting` whenever there is data to judge (we never invent the 50% floor). With zero
+ * serves it is `unmeasured` (nothing served yet). The named cause for a high no-domain share is
+ * "domains unratified".
+ */
+export function diagnoseM1(m1) {
+    const reading = `served=${m1.mix.total} · audited=${pct(m1.auditedShare)} · draft=${pct(m1.draftShare)} · ` +
+        `no-domain=${pct(m1.noDomainShare)} (window ${ms(m1.windowMs)})`;
+    if (m1.mix.total === 0) {
+        // TEN-2449: M1 counts ONE serve shape (`orient/cli/brief`). Naming start_pb here would loop
+        // the operator forever: that surface is uninstrumented, so serving it can never move M1.
+        const outOfCoverage = m1.excludedOutOfCoverage ?? 0;
+        return {
+            metric: 'M1',
+            label: 'Retrieval truth-state mix',
+            status: 'unmeasured',
+            reading,
+            cause: outOfCoverage > 0
+                ? `${outOfCoverage} row(s) served but outside the covered path (orient/cli/brief) — nothing counted`
+                : 'no governance served in the window yet — nothing to classify',
+            nextAction: 'run `pb orient -b` in a terminal (the covered orient/cli/brief path), then re-read',
+        };
+    }
+    return {
+        metric: 'M1',
+        label: 'Retrieval truth-state mix',
+        status: 'awaiting',
+        reading,
+        cause: `no-domain-share floor ${UNPINNED}`,
+        // INS-1775 records the ≈57% no-domain baseline — cited here in a comment, not the user string.
+        nextAction: 'pin the no-domain floor (≈57% measured baseline); if no-domain stays high, domains are unratified → ratify the inferred-domain binding entries',
+    };
+}
+/**
+ * M2 — ranking agreement (tests the unify integrity). This one HAS a clean structural signal:
+ * post-unify, full parity (divergentBuckets === 0) is the EXPECTED steady state — that's `ok`.
+ * A non-zero divergence IS the regression signal — `degraded` — and the M2-magnitude flip stays
+ * unpinned, but the structural ok/degraded does not need a number (0 vs non-0 is exact).
+ */
+export function diagnoseM2(m2) {
+    const reading = `divergentBuckets=${m2.divergentBuckets} · fullParity=${m2.fullParity}${m2.task ? ` · task="${m2.task}"` : ''}`;
+    if (m2.fullParity && m2.divergentBuckets === 0) {
+        return {
+            metric: 'M2',
+            label: 'Ranking agreement (unify integrity)',
+            status: 'ok',
+            reading,
+            cause: 'orient ↔ start_pb render identical set+order+tier — the unify holds',
+            nextAction: '',
+        };
+    }
+    return {
+        metric: 'M2',
+        label: 'Ranking agreement (unify integrity)',
+        status: 'degraded',
+        reading,
+        cause: 'the two surfaces re-forked — one is selecting/ordering governance outside the shared ranker',
+        // The ranking-unify tension is TEN-2431 — cited in the comment, kept out of the user string.
+        nextAction: 'diff getOrientEntries vs getBindingGovernanceView; restore the shared ranker as the single source',
+    };
+}
+/**
+ * M3 — draft-to-SSOT rate. The in-session-ratify floor (ASM-45 rolling-20 window) is UNPINNED,
+ * so a judged reading is `awaiting`. Zero captures in the window → `unmeasured`.
+ */
+export function diagnoseM3(m3) {
+    const reading = `captured=${m3.capturedCount} · in-session=${pct(m3.inSessionRatifiedRate)} · ` +
+        `ever=${pct(m3.everRatifiedRate)} (over ${m3.sessionsConsidered}/${m3.sessionWindow} sessions)`;
+    if (m3.capturedCount === 0) {
+        return {
+            metric: 'M3',
+            label: 'Draft-to-SSOT rate',
+            status: 'unmeasured',
+            reading,
+            cause: 'no captures in the rolling session window — nothing to rate',
+            nextAction: 'capture in-session (pb capture), then re-read after the human accepts in Cortex',
+        };
+    }
+    return {
+        metric: 'M3',
+        label: 'Draft-to-SSOT rate',
+        status: 'awaiting',
+        reading,
+        // ASM-45 (the draft-graveyard hypothesis + rolling-20-session design) governs this metric —
+        // cited in comments; the user strings stay ID-free.
+        cause: `in-session ratify floor ${UNPINNED} (rolling-20-session design) — PROPOSED pin ${pct(M3_PROPOSED_IN_SESSION_FLOOR)}, awaiting founder accept (Product-OS calibration baseline, not this workspace's own history)`,
+        // Round 1 scoped `cause` but left `nextAction` — the formatter prints this verbatim too.
+        nextAction: `pin the in-session floor (Product-OS calibration proposal ${pct(M3_PROPOSED_IN_SESSION_FLOOR)}, not derived from this workspace, awaiting founder accept); if the rate stays below it, the draft-graveyard pattern is confirmed → accept/promote captures in Cortex`,
+    };
+}
+/**
+ * M4 — organic rework meter (WP-487 Slice 2, FEAT-1374). Joins served governance against
+ * later-reported rework signals within a frozen 72h window (TEN-2728) and reports RAW
+ * distribution + a false-attribution counter-metric — never a verdict (DEC-1282 unchanged for
+ * M4) and never a causality claim (spec §6: this is association, not "context served CAUSED
+ * first-time-good"). `unmeasured` when nothing has been served in the window yet (no coverage
+ * population to judge, mirrors diagnoseM1's zero-mix gate); `awaiting` whenever there IS data,
+ * since no first-time-good floor is pinned yet — the raw share is always shown, never replaced
+ * by an invented threshold.
+ */
+export function diagnoseM4(m4) {
+    const reading = `served=${m4.servedCount} (decided=${m4.decidedCount}, pending=${m4.pendingObservationCount}, ` +
+        `unattributable=${m4.unattributableServedCount}) · first-time-good=${pct(m4.firstTimeGoodShare)} ` +
+        `(n=${m4.decidedCount}) · reworked=${m4.reworkedCount} · signals=${m4.signalCount} ` +
+        `(revision=${m4.byClassification.revision}, ci-review=${m4.byClassification.ci_review_correction}, ` +
+        `override=${m4.byClassification.user_override}) · false-attribution=${pct(m4.falseAttributionRate)}`;
+    if (m4.servedCount === 0) {
+        return {
+            metric: 'M4',
+            label: 'Organic rework meter (first-time-good)',
+            status: 'unmeasured',
+            reading,
+            cause: 'no governance served in the window yet — nothing to join rework signals against',
+            nextAction: 'run `pb orient -b` to serve governance, then report rework via `pb rework report` as it occurs',
+        };
+    }
+    return {
+        metric: 'M4',
+        label: 'Organic rework meter (first-time-good)',
+        status: 'awaiting',
+        reading,
+        // No first-time-good floor is pinned yet (this WP's job was to make a baseline possible, not
+        // to claim one — WP-C spec §5/§9). No causality claim: a served row followed by rework within
+        // the window is an ASSOCIATION, never asserted as "the served context caused the rework".
+        cause: `first-time-good floor ${UNPINNED}; a high false-attribution rate means many rework signals have no served-context evidence to join against — read the share with that caveat, not as a verdict`,
+        nextAction: 'pin the first-time-good floor once a predeclared observation window completes (WP-C spec §9); if false-attribution stays high, the pb-reported signal may be missing real rework events a CI/git signal would catch (spec §6 tripwire)',
+    };
+}
+/**
+ * M5 — projection staleness, computed as LOCAL-PULL drift at the CLI (DEC-1284). The staleness-age
+ * flip is UNPINNED, so a determinable reading is `awaiting`. When the local `.productbrain/`
+ * projection (or the server materialization) is unavailable, the status is `unmeasured` with the
+ * reason — fail-soft, never a fake ok.
+ */
+export function diagnoseM5(m5) {
+    if (m5.unavailableReason || m5.localProjectionMtime === null) {
+        return {
+            metric: 'M5',
+            label: 'Projection staleness (local-pull drift)',
+            status: 'unmeasured',
+            reading: m5.unavailableReason ?? 'local .productbrain/ projection not found',
+            cause: 'cannot stat the local projection — local-pull drift is undeterminable here',
+            nextAction: 'run from a worktree with a materialized .productbrain/ (pb handshake), then re-read',
+        };
+    }
+    // The local projection exists, but with NO server materialization (e.g. missing/expired
+    // setup_receipt) there is no reference to diff against — the local-pull DRIFT (server−local,
+    // this metric's whole definition) is undeterminable. Report `unmeasured` (honesty over flattery)
+    // rather than `awaiting` with a `local-lag=n/a` that implies the drift is merely awaiting a
+    // threshold. local-age is still surfaced as the one signal we can honestly read.
+    if (m5.serverMaterializedAt === null || m5.localLagMs === null) {
+        return {
+            metric: 'M5',
+            label: 'Projection staleness (local-pull drift)',
+            status: 'unmeasured',
+            reading: `local-age=${ms(m5.localAgeMs)} · server materialization: none (drift undeterminable)`,
+            cause: 'no server materialization to diff against — local-pull drift (server−local) cannot be computed',
+            nextAction: 'materialize the projection server-side (pb handshake), then re-read to measure drift',
+        };
+    }
+    const lag = m5.localLagMs;
+    const age = m5.localAgeMs;
+    const reading = `local-lag=${ms(lag)} (server−local) · local-age=${ms(age)}`;
+    return {
+        metric: 'M5',
+        label: 'Projection staleness (local-pull drift)',
+        status: 'awaiting',
+        reading,
+        cause: `staleness-age threshold ${UNPINNED} — PROPOSED pin ${M5_PROPOSED_LOCAL_AGE_FLOOR_DAYS}d, awaiting founder accept (Product-OS calibration: one on-record local-age reading, not a trend); this is the one metric where PB can be staler than the never-drifting CLAUDE.md file`,
+        nextAction: 're-run pb handshake to re-pull the projection if the lag/age looks high',
+    };
+}
+/**
+ * M5 (server view) — the Cortex page has NO local filesystem, so it cannot compute the spec's
+ * local-pull drift (DEC-1284). It reads only the server-side projection-generation building block
+ * (`materialization.latestMaterializedAt` / age) the public action already returns — adding NO new
+ * truth (spec §2d). This is deliberately a DIFFERENT, HONEST framing from `diagnoseM5`: it is
+ * SERVER-generation age, never claimed to be the local-pull staleness the CLI measures. The
+ * staleness-age flip is UNPINNED → `awaiting` when there is a materialization; `unmeasured` when
+ * the workspace has never materialized.
+ */
+export function diagnoseM5FromServer(materialization) {
+    if (materialization.latestMaterializedAt === null) {
+        return {
+            metric: 'M5',
+            label: 'Projection staleness (server generation age)',
+            status: 'unmeasured',
+            reading: 'no materialization receipt yet — the projection has never been generated server-side',
+            cause: 'nothing to age — the workspace has not materialized its projection',
+            nextAction: 'run pb handshake to materialize the projection, then re-read',
+        };
+    }
+    const age = materialization.materializationAgeMs;
+    return {
+        metric: 'M5',
+        label: 'Projection staleness (server generation age)',
+        status: 'awaiting',
+        reading: `server-generation age=${ms(age)} (browser view: local-pull drift is CLI-only)`,
+        cause: `staleness-age threshold ${UNPINNED} — PROPOSED pin ${M5_PROPOSED_SERVER_AGE_FLOOR_DAYS}d, awaiting founder accept (Product-OS calibration: one on-record server-generation-age reading, not a trend); this is the one metric where PB can be staler than the never-drifting CLAUDE.md file`,
+        nextAction: 're-run pb handshake to re-generate the projection if the age looks high',
+    };
+}
+/**
+ * The STANDING coaching/moat-quality `unmeasured` line (TEN-2377). It is NOT a tile with a number —
+ * it is a permanent honesty marker so four green tiles are never read as "the moat is healthy".
+ */
+export function diagnoseCoachingMoat() {
+    return {
+        metric: 'MOAT',
+        label: 'Coaching / moat quality',
+        status: 'unmeasured',
+        // The standing-unmeasured tension is TEN-2377; the real instrument is ASM-46's
+        // blind-padding-rate test. Both cited in comments, kept out of the user-facing strings.
+        reading: 'structurally unmeasured',
+        cause: 'the coach skips decision/standard/business-rule (the WHY-moat types) — a counter would flatter the board',
+        nextAction: 'do NOT read four green tiles as a healthy moat; the real instrument is the blind-padding-rate test',
+    };
+}
+/**
+ * diagnose — the whole mapping: data-layer payload + CLI-computed M5 local-drift → per-metric
+ * diagnoses + the standing M4-awaiting + coaching-unmeasured lines. Pure.
+ */
+export function diagnose(payload, m5) {
+    return {
+        m1: diagnoseM1(payload.m1),
+        m2: diagnoseM2(payload.m2),
+        m3: diagnoseM3(payload.m3),
+        m4: diagnoseM4(payload.m4),
+        m5: diagnoseM5(m5),
+        moat: diagnoseCoachingMoat(),
+    };
+}
+/**
+ * diagnoseFromServer — the Cortex-page mapping: it has the public action payload but NO local fs,
+ * so M5 is the server-generation-age view (`diagnoseM5FromServer`) rather than the CLI local-pull
+ * drift. Everything else is identical to `diagnose`. Pure, browser-safe.
+ */
+export function diagnoseFromServer(payload) {
+    return {
+        m1: diagnoseM1(payload.m1),
+        m2: diagnoseM2(payload.m2),
+        m3: diagnoseM3(payload.m3),
+        m4: diagnoseM4(payload.m4),
+        m5: diagnoseM5FromServer(payload.materialization),
+        moat: diagnoseCoachingMoat(),
+    };
+}
+//# sourceMappingURL=diagnose.js.map

package/dist/scoreboard/diagnose.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"diagnose.js","sourceRoot":"","sources":["../../src/scoreboard/diagnose.ts"],"names":[],"mappings":"AAAA;;;;;;;GAOG;AAoLH,4FAA4F;AAC5F,8FAA8F;AAC9F,iGAAiG;AACjG,uDAAuD;AACvD,MAAM,QAAQ,GAAG,wDAAwD,CAAC;AAE1E,yFAAyF;AACzF,mGAAmG;AACnG,oGAAoG;AACpG,2FAA2F;AAC3F,kGAAkG;AAClG,wFAAwF;AACxF,2FAA2F;AAC3F,8FAA8F;AAC9F,2FAA2F;AAC3F,+FAA+F;AAC/F,MAAM,4BAA4B,GAAG,GAAG,CAAC;AACzC,mGAAmG;AACnG,2FAA2F;AAC3F,kGAAkG;AAClG,mGAAmG;AACnG,mGAAmG;AACnG,MAAM,gCAAgC,GAAG,EAAE,CAAC;AAC5C,MAAM,iCAAiC,GAAG,CAAC,CAAC;AAE5C,SAAS,GAAG,CAAC,CAAgB;IAC3B,OAAO,CAAC,KAAK,IAAI,CAAC,CAAC,CAAC,KAAK,CAAC,CAAC,CAAC,GAAG,CAAC,CAAC,GAAG,GAAG,CAAC,CAAC,OAAO,CAAC,CAAC,CAAC,GAAG,CAAC;AACzD,CAAC;AAED,SAAS,EAAE,CAAC,CAAgB;IAC1B,IAAI,CAAC,KAAK,IAAI;QAAE,OAAO,KAAK,CAAC;IAC7B,MAAM,CAAC,GAAG,CAAC,GAAG,CAAC,IAAI,GAAG,EAAE,GAAG,EAAE,CAAC,CAAC;IAC/B,IAAI,CAAC,GAAG,CAAC;QAAE,OAAO,GAAG,CAAC,CAAC,GAAG,CAAC,IAAI,GAAG,EAAE,CAAC,CAAC,CAAC,OAAO,CAAC,CAAC,CAAC,GAAG,CAAC;IACrD,IAAI,CAAC,GAAG,EAAE;QAAE,OAAO,GAAG,CAAC,CAAC,OAAO,CAAC,CAAC,CAAC,GAAG,CAAC;IACtC,OAAO,GAAG,CAAC,CAAC,GAAG,EAAE,CAAC,CAAC,OAAO,CAAC,CAAC,CAAC,GAAG,CAAC;AACnC,CAAC;AAED;;;;;GAKG;AACH,MAAM,UAAU,UAAU,CAAC,EAAa;IACtC,MAAM,OAAO,GACX,UAAU,EAAE,CAAC,GAAG,CAAC,KAAK,cAAc,GAAG,CAAC,EAAE,CAAC,YAAY,CAAC,YAAY,GAAG,CAAC,EAAE,CAAC,UAAU,CAAC,KAAK;QAC3F,aAAa,GAAG,CAAC,EAAE,CAAC,aAAa,CAAC,YAAY,EAAE,CAAC,EAAE,CAAC,QAAQ,CAAC,GAAG,CAAC;IACnE,IAAI,EAAE,CAAC,GAAG,CAAC,KAAK,KAAK,CAAC,EAAE,CAAC;QACvB,4FAA4F;QAC5F,yFAAyF;QACzF,MAAM,aAAa,GAAG,EAAE,CAAC,qBAAqB,IAAI,CAAC,CAAC;QACpD,OAAO;YACL,MAAM,EAAE,IAAI;YACZ,KAAK,EAAE,2BAA2B;YAClC,MAAM,EAAE,YAAY;YACpB,OAAO;YACP,KAAK,EACH,aAAa,GAAG,CAAC;gBACf,CAAC,CAAC,GAAG,aAAa,kFAAkF;gBACpG,CAAC,CAAC,8DAA8D;YACpE,UAAU,EAAE,oFAAoF;SACjG,CAAC;IACJ,CAAC;IACD,OAAO;QACL,MAAM,EAAE,IAAI;QACZ,KAAK,EAAE,2BAA2B;QAClC,MAAM,EAAE,UAAU;QAClB,OAAO;QACP,KAAK,EAAE,yBAAyB,QAAQ,EAAE;QAC1C,+FAA+F;QAC/F,UAAU,EACR,gJAAgJ;KACnJ,CAAC;AACJ,CAAC;AAED;;;;;GAKG;AACH,MAAM,UAAU,UAAU,CAAC,EAAa;IACtC,MAAM,OAAO,GAAG,oBAAoB,EAAE,CAAC,gBAAgB,iBAAiB,EAAE,CAAC,UAAU,GAAG,EAAE,CAAC,IAAI,CAAC,CAAC,CAAC,YAAY,EAAE,CAAC,IAAI,GAAG,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC;IAChI,IAAI,EAAE,CAAC,UAAU,IAAI,EAAE,CAAC,gBAAgB,KAAK,CAAC,EAAE,CAAC;QAC/C,OAAO;YACL,MAAM,EAAE,IAAI;YACZ,KAAK,EAAE,qCAAqC;YAC5C,MAAM,EAAE,IAAI;YACZ,OAAO;YACP,KAAK,EAAE,qEAAqE;YAC5E,UAAU,EAAE,EAAE;SACf,CAAC;IACJ,CAAC;IACD,OAAO;QACL,MAAM,EAAE,IAAI;QACZ,KAAK,EAAE,qCAAqC;QAC5C,MAAM,EAAE,UAAU;QAClB,OAAO;QACP,KAAK,EAAE,6FAA6F;QACpG,6FAA6F;QAC7F,UAAU,EACR,mGAAmG;KACtG,CAAC;AACJ,CAAC;AAED;;;GAGG;AACH,MAAM,UAAU,UAAU,CAAC,EAAa;IACtC,MAAM,OAAO,GACX,YAAY,EAAE,CAAC,aAAa,iBAAiB,GAAG,CAAC,EAAE,CAAC,qBAAqB,CAAC,KAAK;QAC/E,QAAQ,GAAG,CAAC,EAAE,CAAC,gBAAgB,CAAC,UAAU,EAAE,CAAC,kBAAkB,IAAI,EAAE,CAAC,aAAa,YAAY,CAAC;IAClG,IAAI,EAAE,CAAC,aAAa,KAAK,CAAC,EAAE,CAAC;QAC3B,OAAO;YACL,MAAM,EAAE,IAAI;YACZ,KAAK,EAAE,oBAAoB;YAC3B,MAAM,EAAE,YAAY;YACpB,OAAO;YACP,KAAK,EAAE,6DAA6D;YACpE,UAAU,EAAE,iFAAiF;SAC9F,CAAC;IACJ,CAAC;IACD,OAAO;QACL,MAAM,EAAE,IAAI;QACZ,KAAK,EAAE,oBAAoB;QAC3B,MAAM,EAAE,UAAU;QAClB,OAAO;QACP,4FAA4F;QAC5F,oDAAoD;QACpD,KAAK,EAAE,2BAA2B,QAAQ,+CAA+C,GAAG,CAAC,4BAA4B,CAAC,+FAA+F;QACzN,yFAAyF;QACzF,UAAU,EACR,6DAA6D,GAAG,CAAC,4BAA4B,CAAC,uKAAuK;KACxQ,CAAC;AACJ,CAAC;AAED;;;;;;;;;GASG;AACH,MAAM,UAAU,UAAU,CAAC,EAAa;IACtC,MAAM,OAAO,GACX,UAAU,EAAE,CAAC,WAAW,aAAa,EAAE,CAAC,YAAY,aAAa,EAAE,CAAC,uBAAuB,IAAI;QAC/F,kBAAkB,EAAE,CAAC,yBAAyB,uBAAuB,GAAG,CAAC,EAAE,CAAC,kBAAkB,CAAC,GAAG;QAClG,MAAM,EAAE,CAAC,YAAY,gBAAgB,EAAE,CAAC,aAAa,cAAc,EAAE,CAAC,WAAW,GAAG;QACpF,aAAa,EAAE,CAAC,gBAAgB,CAAC,QAAQ,eAAe,EAAE,CAAC,gBAAgB,CAAC,oBAAoB,IAAI;QACpG,YAAY,EAAE,CAAC,gBAAgB,CAAC,aAAa,yBAAyB,GAAG,CAAC,EAAE,CAAC,oBAAoB,CAAC,EAAE,CAAC;IACvG,IAAI,EAAE,CAAC,WAAW,KAAK,CAAC,EAAE,CAAC;QACzB,OAAO;YACL,MAAM,EAAE,IAAI;YACZ,KAAK,EAAE,wCAAwC;YAC/C,MAAM,EAAE,YAAY;YACpB,OAAO;YACP,KAAK,EAAE,iFAAiF;YACxF,UAAU,EAAE,gGAAgG;SAC7G,CAAC;IACJ,CAAC;IACD,OAAO;QACL,MAAM,EAAE,IAAI;QACZ,KAAK,EAAE,wCAAwC;QAC/C,MAAM,EAAE,UAAU;QAClB,OAAO;QACP,6FAA6F;QAC7F,8FAA8F;QAC9F,0FAA0F;QAC1F,KAAK,EAAE,yBAAyB,QAAQ,+JAA+J;QACvM,UAAU,EACR,wOAAwO;KAC3O,CAAC;AACJ,CAAC;AAED;;;;;GAKG;AACH,MAAM,UAAU,UAAU,CAAC,EAAgB;IACzC,IAAI,EAAE,CAAC,iBAAiB,IAAI,EAAE,CAAC,oBAAoB,KAAK,IAAI,EAAE,CAAC;QAC7D,OAAO;YACL,MAAM,EAAE,IAAI;YACZ,KAAK,EAAE,yCAAyC;YAChD,MAAM,EAAE,YAAY;YACpB,OAAO,EAAE,EAAE,CAAC,iBAAiB,IAAI,2CAA2C;YAC5E,KAAK,EAAE,4EAA4E;YACnF,UAAU,EAAE,qFAAqF;SAClG,CAAC;IACJ,CAAC;IACD,wFAAwF;IACxF,6FAA6F;IAC7F,iGAAiG;IACjG,4FAA4F;IAC5F,iFAAiF;IACjF,IAAI,EAAE,CAAC,oBAAoB,KAAK,IAAI,IAAI,EAAE,CAAC,UAAU,KAAK,IAAI,EAAE,CAAC;QAC/D,OAAO;YACL,MAAM,EAAE,IAAI;YACZ,KAAK,EAAE,yCAAyC;YAChD,MAAM,EAAE,YAAY;YACpB,OAAO,EAAE,aAAa,EAAE,CAAC,EAAE,CAAC,UAAU,CAAC,wDAAwD;YAC/F,KAAK,EAAE,gGAAgG;YACvG,UAAU,EAAE,sFAAsF;SACnG,CAAC;IACJ,CAAC;IACD,MAAM,GAAG,GAAG,EAAE,CAAC,UAAU,CAAC;IAC1B,MAAM,GAAG,GAAG,EAAE,CAAC,UAAU,CAAC;IAC1B,MAAM,OAAO,GAAG,aAAa,EAAE,CAAC,GAAG,CAAC,+BAA+B,EAAE,CAAC,GAAG,CAAC,EAAE,CAAC;IAC7E,OAAO;QACL,MAAM,EAAE,IAAI;QACZ,KAAK,EAAE,yCAAyC;QAChD,MAAM,EAAE,UAAU;QAClB,OAAO;QACP,KAAK,EAAE,2BAA2B,QAAQ,mBAAmB,gCAAgC,yLAAyL;QACtR,UAAU,EAAE,yEAAyE;KACtF,CAAC;AACJ,CAAC;AAED;;;;;;;;GAQG;AACH,MAAM,UAAU,oBAAoB,CAAC,eAAuC;IAC1E,IAAI,eAAe,CAAC,oBAAoB,KAAK,IAAI,EAAE,CAAC;QAClD,OAAO;YACL,MAAM,EAAE,IAAI;YACZ,KAAK,EAAE,8CAA8C;YACrD,MAAM,EAAE,YAAY;YACpB,OAAO,EAAE,sFAAsF;YAC/F,KAAK,EAAE,oEAAoE;YAC3E,UAAU,EAAE,8DAA8D;SAC3E,CAAC;IACJ,CAAC;IACD,MAAM,GAAG,GAAG,eAAe,CAAC,oBAAoB,CAAC;IACjD,OAAO;QACL,MAAM,EAAE,IAAI;QACZ,KAAK,EAAE,8CAA8C;QACrD,MAAM,EAAE,UAAU;QAClB,OAAO,EACL,yBAAyB,EAAE,CAAC,GAAG,CAAC,+CAA+C;QACjF,KAAK,EAAE,2BAA2B,QAAQ,mBAAmB,iCAAiC,qMAAqM;QACnS,UAAU,EAAE,yEAAyE;KACtF,CAAC;AACJ,CAAC;AAED;;;GAGG;AACH,MAAM,UAAU,oBAAoB;IAClC,OAAO;QACL,MAAM,EAAE,MAAM;QACd,KAAK,EAAE,yBAAyB;QAChC,MAAM,EAAE,YAAY;QACpB,+EAA+E;QAC/E,wFAAwF;QACxF,OAAO,EAAE,yBAAyB;QAClC,KAAK,EAAE,0GAA0G;QACjH,UAAU,EAAE,oGAAoG;KACjH,CAAC;AACJ,CAAC;AAaD;;;GAGG;AACH,MAAM,UAAU,QAAQ,CAAC,OAA0B,EAAE,EAAgB;IACnE,OAAO;QACL,EAAE,EAAE,UAAU,CAAC,OAAO,CAAC,EAAE,CAAC;QAC1B,EAAE,EAAE,UAAU,CAAC,OAAO,CAAC,EAAE,CAAC;QAC1B,EAAE,EAAE,UAAU,CAAC,OAAO,CAAC,EAAE,CAAC;QAC1B,EAAE,EAAE,UAAU,CAAC,OAAO,CAAC,EAAE,CAAC;QAC1B,EAAE,EAAE,UAAU,CAAC,EAAE,CAAC;QAClB,IAAI,EAAE,oBAAoB,EAAE;KAC7B,CAAC;AACJ,CAAC;AAED;;;;GAIG;AACH,MAAM,UAAU,kBAAkB,CAAC,OAA0B;IAC3D,OAAO;QACL,EAAE,EAAE,UAAU,CAAC,OAAO,CAAC,EAAE,CAAC;QAC1B,EAAE,EAAE,UAAU,CAAC,OAAO,CAAC,EAAE,CAAC;QAC1B,EAAE,EAAE,UAAU,CAAC,OAAO,CAAC,EAAE,CAAC;QAC1B,EAAE,EAAE,UAAU,CAAC,OAAO,CAAC,EAAE,CAAC;QAC1B,EAAE,EAAE,oBAAoB,CAAC,OAAO,CAAC,eAAe,CAAC;QACjD,IAAI,EAAE,oBAAoB,EAAE;KAC7B,CAAC;AACJ,CAAC"}

package/dist/scoreboard/diagnose.test.d.ts

@@ -0,0 +1,14 @@
+/**
+ * Diagnosis-layer unit tests (spec §2d, A4).
+ *
+ * Asserts the pure metric→{status,reading,cause,nextAction} mapping, with explicit coverage of
+ * the awaiting / unmeasured / degraded / ok states and the honesty invariants:
+ *   - No flip threshold is pinned (DEC-1282) → judged readings are `awaiting`, never an invented
+ *     number, and the raw reading is always present.
+ *   - M4 (WP-487 Slice 2, FEAT-1374: the organic rework meter) reports raw distribution only —
+ *     `unmeasured` with nothing served yet, `awaiting` otherwise, never a fake green, never a
+ *     causality claim.
+ *   - Coaching/moat quality is a standing `unmeasured` line (TEN-2377).
+ */
+export {};
+//# sourceMappingURL=diagnose.test.d.ts.map

package/dist/scoreboard/diagnose.test.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"diagnose.test.d.ts","sourceRoot":"","sources":["../../src/scoreboard/diagnose.test.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;GAWG"}

package/dist/scoreboard/diagnose.test.js

@@ -0,0 +1,311 @@
+/**
+ * Diagnosis-layer unit tests (spec §2d, A4).
+ *
+ * Asserts the pure metric→{status,reading,cause,nextAction} mapping, with explicit coverage of
+ * the awaiting / unmeasured / degraded / ok states and the honesty invariants:
+ *   - No flip threshold is pinned (DEC-1282) → judged readings are `awaiting`, never an invented
+ *     number, and the raw reading is always present.
+ *   - M4 (WP-487 Slice 2, FEAT-1374: the organic rework meter) reports raw distribution only —
+ *     `unmeasured` with nothing served yet, `awaiting` otherwise, never a fake green, never a
+ *     causality claim.
+ *   - Coaching/moat quality is a standing `unmeasured` line (TEN-2377).
+ */
+import { describe, it, expect } from 'vitest';
+import { diagnose, diagnoseM1, diagnoseM2, diagnoseM3, diagnoseM4, diagnoseM5, diagnoseM5FromServer, diagnoseCoachingMoat, } from './diagnose.js';
+const m1WithData = {
+    mix: { audited: 3, draft: 2, noDomain: 5, total: 10 },
+    noDomainShare: 0.5,
+    auditedShare: 0.3,
+    draftShare: 0.2,
+    windowMs: 30 * 24 * 60 * 60 * 1000,
+    servedRowCount: 10,
+};
+const m1Empty = {
+    mix: { audited: 0, draft: 0, noDomain: 0, total: 0 },
+    noDomainShare: null,
+    auditedShare: null,
+    draftShare: null,
+    windowMs: 30 * 24 * 60 * 60 * 1000,
+    servedRowCount: 0,
+};
+describe('diagnoseM1', () => {
+    it('is awaiting (never an invented floor) when there is data — DEC-1282', () => {
+        const d = diagnoseM1(m1WithData);
+        expect(d.status).toBe('awaiting');
+        expect(d.metric).toBe('M1');
+        // Raw reading always shown.
+        expect(d.reading).toContain('no-domain=50%');
+        expect(d.reading).toContain('served=10');
+        // Unpinned cause + concrete next action (domains unratified).
+        expect(d.cause).toMatch(/unpinned/i);
+        expect(d.nextAction).toMatch(/ratify/i);
+    });
+    it('is unmeasured when nothing was served', () => {
+        const d = diagnoseM1(m1Empty);
+        expect(d.status).toBe('unmeasured');
+        // TEN-2449: must name the ONE covered path. Recommending start_pb would loop the operator
+        // forever — that surface is uninstrumented, so serving it can never move M1 off `unmeasured`.
+        expect(d.nextAction).toMatch(/orient\/cli\/brief/);
+        expect(d.nextAction).not.toMatch(/start_pb/);
+        expect(d.cause).toMatch(/no governance served/i);
+    });
+    it('unmeasured-but-serving: names the coverage gap instead of repeating "drive a serve"', () => {
+        const d = diagnoseM1({ ...m1Empty, excludedOutOfCoverage: 5 });
+        expect(d.status).toBe('unmeasured');
+        // The operator HAS been serving — just not on the covered path. Saying "nothing served" here
+        // would send them hunting for a write bug that does not exist.
+        expect(d.cause).toMatch(/outside the covered path/i);
+        expect(d.cause).toContain('5 row(s)');
+    });
+});
+describe('diagnoseM2', () => {
+    it('is ok at full parity (post-unify steady state — 0 vs non-0 is exact, no number invented)', () => {
+        const m2 = { divergentBuckets: 0, fullParity: true, task: null };
+        const d = diagnoseM2(m2);
+        expect(d.status).toBe('ok');
+        expect(d.nextAction).toBe('');
+        expect(d.reading).toContain('divergentBuckets=0');
+    });
+    it('is degraded when the surfaces re-fork (regression signal)', () => {
+        const m2 = { divergentBuckets: 2, fullParity: false, task: 'do a thing' };
+        const d = diagnoseM2(m2);
+        expect(d.status).toBe('degraded');
+        expect(d.reading).toContain('divergentBuckets=2');
+        expect(d.reading).toContain('task="do a thing"');
+        expect(d.nextAction).toMatch(/shared ranker|TEN-2431/);
+    });
+});
+describe('diagnoseM3', () => {
+    it('is awaiting when there are captures (floor unpinned)', () => {
+        const m3 = {
+            capturedCount: 8,
+            everRatifiedCount: 3,
+            inSessionRatifiedCount: 1,
+            everRatifiedRate: 0.375,
+            inSessionRatifiedRate: 0.125,
+            sessionWindow: 20,
+            sessionsConsidered: 12,
+        };
+        const d = diagnoseM3(m3);
+        expect(d.status).toBe('awaiting');
+        expect(d.reading).toContain('captured=8');
+        expect(d.cause).toMatch(/unpinned/i);
+        // WP-527 (Lane T Slice 2): a PROPOSED (not pinned) floor value now rides alongside the
+        // unpinned status — the tile must never claim ok/degraded from a proposal (status stays
+        // `awaiting` above), but the candidate number + "awaiting founder accept" must be visible.
+        expect(d.cause).toMatch(/PROPOSED pin 10%/);
+        expect(d.cause).toMatch(/awaiting founder accept/);
+        expect(d.nextAction).toMatch(/10%/);
+        // PR #477 review round 1 (codex P2): this is a SHARED function with no workspaceId — the
+        // calibration number must never read as if it came from the calling workspace's own
+        // history (a different tenant would see the same "every measured run" claim wrongly).
+        expect(d.cause).toMatch(/Product-OS calibration/);
+        expect(d.cause).toMatch(/not this workspace's own history/);
+        // PR #477 review round 2 (codex P2): round 1 only scoped `cause` — `nextAction` must be
+        // scoped too, since the CLI formatter prints it verbatim to every tenant just the same.
+        expect(d.nextAction).toMatch(/not derived from this workspace/);
+        // chain-id-leak guard: the user-facing strings carry no raw Chain IDs.
+        expect(d.cause).not.toMatch(/\b[A-Z]{2,5}-\d+\b/);
+        expect(d.nextAction).not.toMatch(/\b[A-Z]{2,5}-\d+\b/);
+    });
+    it('is unmeasured with zero captures', () => {
+        const m3 = {
+            capturedCount: 0,
+            everRatifiedCount: 0,
+            inSessionRatifiedCount: 0,
+            everRatifiedRate: null,
+            inSessionRatifiedRate: null,
+            sessionWindow: 20,
+            sessionsConsidered: 5,
+        };
+        const d = diagnoseM3(m3);
+        expect(d.status).toBe('unmeasured');
+    });
+});
+const m4Empty = {
+    reworkJoinWindowMs: 72 * 60 * 60 * 1000,
+    lookbackWindowMs: 30 * 24 * 60 * 60 * 1000,
+    now: Date.now(),
+    servedCount: 0,
+    unattributableServedCount: 0,
+    pendingObservationCount: 0,
+    decidedCount: 0,
+    firstTimeGoodCount: 0,
+    reworkedCount: 0,
+    firstTimeGoodShare: null,
+    signalCount: 0,
+    joinedSignalCount: 0,
+    unjoinedSignalCount: 0,
+    falseAttributionRate: null,
+    byClassification: { revision: 0, ci_review_correction: 0, user_override: 0 },
+    ciReviewCorrectionAttribution: { attributable: 0, notAttributable: 0, unspecified: 0 },
+    cohort: { workspaceId: 'ws_1', taskProvidedCount: 0, taskAbsentCount: 0 },
+};
+const m4WithData = {
+    reworkJoinWindowMs: 72 * 60 * 60 * 1000,
+    lookbackWindowMs: 30 * 24 * 60 * 60 * 1000,
+    now: Date.now(),
+    servedCount: 10,
+    unattributableServedCount: 1,
+    pendingObservationCount: 2,
+    decidedCount: 7,
+    firstTimeGoodCount: 5,
+    reworkedCount: 2,
+    firstTimeGoodShare: 5 / 7,
+    signalCount: 3,
+    joinedSignalCount: 2,
+    unjoinedSignalCount: 1,
+    falseAttributionRate: 1 / 3,
+    byClassification: { revision: 2, ci_review_correction: 1, user_override: 0 },
+    ciReviewCorrectionAttribution: { attributable: 1, notAttributable: 0, unspecified: 0 },
+    cohort: { workspaceId: 'ws_1', taskProvidedCount: 2, taskAbsentCount: 1 },
+};
+describe('diagnoseM4', () => {
+    it('is unmeasured when nothing has been served in the window yet', () => {
+        const d = diagnoseM4(m4Empty);
+        expect(d.status).toBe('unmeasured');
+        expect(d.metric).toBe('M4');
+        expect(d.cause).toMatch(/nothing to join/i);
+        expect(d.nextAction).toMatch(/pb orient|pb rework report/);
+    });
+    it('is awaiting (raw distribution, no invented threshold) whenever there IS served data', () => {
+        const d = diagnoseM4(m4WithData);
+        expect(d.status).toBe('awaiting');
+        expect(d.metric).toBe('M4');
+        // The raw numbers are always shown, never replaced by a verdict.
+        expect(d.reading).toContain('first-time-good=71%');
+        expect(d.reading).toContain('false-attribution=33%');
+        expect(d.reading).toContain('served=10');
+        expect(d.reading).toContain('signals=3');
+        // No causality claim (spec §6): association only, never "caused"/"proves"/"confirms".
+        expect(d.cause).not.toMatch(/\bcaused\b|\bproves\b|\bconfirms\b/i);
+        // chain-id-leak guard: the user-facing strings carry no raw Chain IDs.
+        expect(d.cause).not.toMatch(/\b[A-Z]{2,5}-\d+\b/);
+        expect(d.nextAction).not.toMatch(/\b[A-Z]{2,5}-\d+\b/);
+    });
+});
+describe('diagnoseM5', () => {
+    it('is unmeasured (fail-soft) when the local projection is unavailable', () => {
+        const m5 = {
+            localProjectionMtime: null,
+            serverMaterializedAt: 123,
+            localLagMs: null,
+            localAgeMs: null,
+            unavailableReason: 'no .productbrain/ project root found',
+        };
+        const d = diagnoseM5(m5);
+        expect(d.status).toBe('unmeasured');
+        expect(d.reading).toContain('no .productbrain/');
+    });
+    it('is awaiting (threshold unpinned) when local drift is determinable', () => {
+        const m5 = {
+            localProjectionMtime: 1_000,
+            serverMaterializedAt: 2_000,
+            localLagMs: 1_000,
+            localAgeMs: 5_000,
+        };
+        const d = diagnoseM5(m5);
+        expect(d.status).toBe('awaiting');
+        expect(d.cause).toMatch(/unpinned/i);
+        // The CLAUDE.md-can't-drift framing is preserved.
+        expect(d.cause).toMatch(/CLAUDE\.md/);
+        // WP-527 (Lane T Slice 2): PROPOSED (not pinned) floor value, status stays `awaiting`.
+        // 10d, not 7d: this is the LOCAL-pull instrument's own single-observation calibration
+        // (PR #477 review fix, codex P2 — a prior draft mixed local-age and server-generation-age
+        // readings as if they were one series; each instrument now gets its own honest number).
+        expect(d.cause).toMatch(/PROPOSED pin 10d/);
+        expect(d.cause).toMatch(/awaiting founder accept/);
+        expect(d.cause).toMatch(/Product-OS calibration/);
+        expect(d.cause).toMatch(/not a trend/);
+        expect(d.cause).not.toMatch(/\b[A-Z]{2,5}-\d+\b/);
+    });
+    it('is unmeasured (not awaiting) when the local projection exists but there is no server reference', () => {
+        // Local projection present, but no server materialization to diff against → the local-pull
+        // DRIFT (server−local) is undeterminable, so the honest status is unmeasured, not a misleading
+        // `awaiting` with local-lag=n/a. local-age is still surfaced.
+        const m5 = {
+            localProjectionMtime: 5_000,
+            serverMaterializedAt: null,
+            localLagMs: null,
+            localAgeMs: 3_000,
+        };
+        const d = diagnoseM5(m5);
+        expect(d.status).toBe('unmeasured');
+        expect(d.reading).toContain('local-age=');
+        expect(d.reading).toMatch(/drift undeterminable/i);
+    });
+});
+describe('diagnoseM5FromServer', () => {
+    it('is awaiting (threshold unpinned) with its OWN server-generation-age PROPOSED value, distinct from diagnoseM5', () => {
+        // PR #477 review fix (codex P2): diagnoseM5 (local-pull drift) and diagnoseM5FromServer
+        // (server-generation age) are two DIFFERENT instruments — a shared "trend" number between
+        // them was miscalibration. Each carries its own single-observation proposal.
+        const materialization = {
+            latestMaterializedAt: 1_000,
+            materializationAgeMs: 2_000,
+            now: 3_000,
+        };
+        const d = diagnoseM5FromServer(materialization);
+        expect(d.status).toBe('awaiting');
+        expect(d.cause).toMatch(/unpinned/i);
+        expect(d.cause).toMatch(/PROPOSED pin 7d/);
+        expect(d.cause).toMatch(/Product-OS calibration/);
+        expect(d.cause).toMatch(/server-generation-age reading/);
+        expect(d.cause).not.toMatch(/\b[A-Z]{2,5}-\d+\b/);
+    });
+});
+describe('diagnoseCoachingMoat', () => {
+    it('is a standing unmeasured line — four green tiles != healthy moat (TEN-2377)', () => {
+        const d = diagnoseCoachingMoat();
+        expect(d.status).toBe('unmeasured');
+        expect(d.cause).toMatch(/WHY-moat|TEN-2377|skips/i);
+        expect(d.nextAction).toMatch(/four green tiles/i);
+    });
+});
+describe('diagnose (bundle)', () => {
+    it('produces all five metric diagnoses plus the standing moat line', () => {
+        const payload = {
+            workspaceId: 'ws_1',
+            generatedAt: Date.now(),
+            m1: m1WithData,
+            m2: { divergentBuckets: 0, fullParity: true, task: null },
+            m3: {
+                capturedCount: 0,
+                everRatifiedCount: 0,
+                inSessionRatifiedCount: 0,
+                everRatifiedRate: null,
+                inSessionRatifiedRate: null,
+                sessionWindow: 20,
+                sessionsConsidered: 0,
+            },
+            m4: m4WithData,
+            materialization: { latestMaterializedAt: null, materializationAgeMs: null, now: Date.now() },
+            coherence: {
+                openConflictCount: 0,
+                openConflictsTruncated: false,
+                reviewed: 0,
+                real: 0,
+                falseAlarm: 0,
+                reviewedTruncated: false,
+                humanBar: { meets: false, lowerBound: null, reviewed: 0 },
+                detectionCoverage: { bySurface: {}, scanned: 0, truncated: false },
+                contradictionOverrideRate: { overrides: 0, heldBlocks: 0, rate: null, truncated: false },
+            },
+        };
+        const m5 = {
+            localProjectionMtime: null,
+            serverMaterializedAt: null,
+            localLagMs: null,
+            localAgeMs: null,
+            unavailableReason: 'absent',
+        };
+        const bundle = diagnose(payload, m5);
+        expect(bundle.m1.status).toBe('awaiting');
+        expect(bundle.m2.status).toBe('ok');
+        expect(bundle.m3.status).toBe('unmeasured');
+        expect(bundle.m4.status).toBe('awaiting');
+        expect(bundle.m5.status).toBe('unmeasured');
+        expect(bundle.moat.status).toBe('unmeasured');
+    });
+});
+//# sourceMappingURL=diagnose.test.js.map