@event4u/agent-config 16.5.0 → 17.0.0

package/src/config/routing-coverage-seed.json

@@ -10,11 +10,11 @@
     },
     "seed": {
         "rules": 0.9626,
-        "skills": 0.4167
+        "skills": 0.5714
     },
     "seed_provenance": {
         "rules": "103 covered of 107 routed rules on 2026-10-06; was 102 of 106 on 2026-09-13, 101 of 105 on 2026-09-08 and 94 of 105 (tier-1 21/24, tier-2 73/81) on 2026-08-25.",
-        "skills": "125 of 300 skills carry evals/triggers.json on 2026-10-11; was 124 of 299 on 2026-10-10, 123 on 2026-10-09, 119 on 2026-10-07, 118 after the third growth of 2026-10-06, 115 after the second growth that day, 114 earlier that day, 112 on 2026-10-04, 111 on 2026-10-02, 107 on 2026-10-01, 103 earlier that day, 102 on 2026-09-30, 101 on 2026-09-19 and 100 on 2026-08-30.",
+        "skills": "172 of 301 skills carry evals/triggers.json on 2026-10-21; was 169 of 301 on 2026-10-20, 168 of 301 on 2026-10-19, 167 of 301 on 2026-10-18 (second growth that day), 157 of 301 earlier that day, 156 of 300 on 2026-10-17, 151 of 300 on 2026-10-16, 148 of 300 earlier that day, 142 of 300 on 2026-10-14, 141 of 300 earlier that day, 140 of 300 earlier still, 137 of 300 earlier still, 134 of 300 earlier still, 127 of 300 on 2026-10-12, 125 of 300 on 2026-10-11, 124 of 299 on 2026-10-10, 123 on 2026-10-09, 119 on 2026-10-07, 118 after the third growth of 2026-10-06, 115 after the second growth that day, 114 earlier that day, 112 on 2026-10-04, 111 on 2026-10-02, 107 on 2026-10-01, 103 earlier that day, 102 on 2026-09-30, 101 on 2026-09-19 and 100 on 2026-08-30.",
         "finding": "The gap between the two seeds IS the roadmap defect D1, quantified. The rules surface is ~90% covered by a deterministic corpus that can fail a PR; the skills surface -- the one production routes on -- is far lower and is covered only by a harness that is advisory only, never gating (rule_trigger_eval.ts:4). The figure is deliberately not restated here (corrected 2026-10-01): it read `25%` against a seed that had since moved to 0.3445, which is the same count-drift this file's own history entries keep repairing one field away. The live value is `seed.skills`, two keys above."
     },
     "history": [
@@ -143,6 +143,97 @@
             "from": 0.4147,
             "to": 0.4167,
             "reason": "`svg-authoring` landed as a new skill on the design family's vector side and shipped its own evals/triggers.json in the same change (124 of 299 -> 125 of 300). This is the FIRST entry on the skills scope where the DENOMINATOR moves: every entry above adds a case to a fixed 299, and this one adds a unit and a case together. The _comment names that as the ratio's other failure direction -- units added without cases -- and the repair it prescribes is the corpus file, not the seed; the file landed with the skill, so the ratio ROSE across a denominator increase (0.4147 -> 0.4167) rather than falling, which is the same shape as the 2026-09-13 and 2026-10-06 entries on the RULES scope. The seed rises to the live value; it may never be LOWERED. The partition is TRAIN, not holdout: sha256('svg-authoring')[0:2] is 0x77 = 119, above the holdout ceiling of 51, so BOTH published train-side measurements are re-taken here on the terms the 2026-10-04, 2026-10-09 and 2026-10-10 entries set. routing-body-signal-verdict.json: verdict unchanged in kind, still `harmful`, false activation +7.50 pp against the +2.0 pp guard, delta-recall 2.360 -> 3.383 pp (still below the 5.0 pp bar), train corpora 97 -> 98, catalogue 299 -> 300. The holdout pin IS touched, for two reasons rather than one: svg-authoring adds a train row, and the same branch EDITED image-provider-routing's existing train corpus -- it splits the vector job shapes between a provider and hand-authoring -- which moves that row's hash while moving no count at all, the case a reader counting files would miss. The sealed holdout set is the same 27 rows and byte-identical, so the ordering claim is untouched. NO entry was removed from trigger_eval_grandfather.json: the skill is new, so it was never allowlisted, and check_trigger_eval_presence fires only on an entry whose skill HAS a corpus."
+        },
+        {
+            "date": "2026-10-12",
+            "scope": "skills",
+            "from": 0.4167,
+            "to": 0.4233,
+            "reason": "road-to-account-lifecycle-threat-rows Phase 1 edited six skill bodies to stop the suite teaching four account-flow anti-patterns, and check_routing_coverage requires a corpus for every skill a diff touches. Two had none: `form-handler` and `judge-security-auditor` (125 -> 127 of 300). Same shape as every entry above and for the same reason: the corpus file is the gate's own remedy rather than a coverage push. Ten cases each -- five exemplars in the explicit and the symptom-first phrasing each skill's own description names, three near-misses naming the neighbour that actually owns the query (fe-design and laravel-validation for a form, threat-modeling and authz-review for a security judge), and two counterexamples guarding each homonym (a Google Form and a form letter; a building badge and a security deposit). The seed rises to the live value; it may never be LOWERED. The partition is TRAIN for both, not holdout: sha256('form-handler')[0:2] is 0x5d = 93 and sha256('judge-security-auditor')[0:2] is 0xc8 = 200, both above the holdout ceiling of 51, so BOTH published train-side measurements are re-taken here on the terms the 2026-10-04, 2026-10-09, 2026-10-10 and 2026-10-11 entries set. routing-body-signal-verdict.json: verdict unchanged in kind, still `harmful`, false activation +7.13 pp against the +2.0 pp guard, delta-recall 3.383 -> 3.71 pp (still below the 5.0 pp bar), train corpora 98 -> 100, catalogue unchanged at 300. The holdout pin IS touched for the row-count reason only -- two train rows added, no holdout row's content changed -- so the sealed 27 rows are byte-identical and the ordering claim is untouched. NO entry was removed from trigger_eval_grandfather.json: neither skill was allowlisted, because neither had a corpus for check_trigger_eval_presence to fire on. CORRECTED 2026-10-13 by drain/local-gates-wired-20261013, and the correction is left beside the original rather than overwriting it, because the error is the instructive part. BOTH SKILLS WERE ALLOWLISTED: `form-handler` and `judge-security-auditor` were in the `skills` array of trigger_eval_grandfather.json at that date and remained there, and giving them a corpus in that change is precisely what made the pair redundant. check_trigger_eval_presence consequently exited 1 from that commit until this one, which removes the two entries. The sentence above reads as a deliberate no-op and was a missed downstream change, and the ratchet went red on the trunk for a day with nothing remote to report it -- itself an argument for the drain this correction arrives in, since that gate ran in `task ci` and in no workflow. The easy way to reproduce the original mistake is worth recording too: the file has four top-level keys and the allowlist is `skills`, so a reader taking the first array in it reads `_no_obligation`, which holds one unrelated entry and answers False for both names."
+        },
+        {
+            "date": "2026-10-14",
+            "scope": "skills",
+            "from": 0.4233,
+            "to": 0.4467,
+            "reason": "road-to-skill-scripts-that-run-where-they-land moved the taught invocation line in seven skill bodies to `node <skills-root>/…/<tool>.mjs`, and check_routing_coverage requires a corpus for every skill a diff touches. None of the seven had one: `api-endpoint`, `blade-ui`, `dashboard-design`, `flux`, `livewire`, `react-shadcn-ui` and `sql-writing` (127 -> 134 of 300). Same shape as every entry above and for the same reason: the corpus file is the gate's own remedy rather than a coverage push. ELEVEN cases each, not the ten every earlier entry records -- six exemplars of which one is German, three near-misses naming the neighbour that actually owns the query, and two counterexamples guarding each homonym; 77 cases in all. The seed rises to the live value; it may never be LOWERED. THE PARTITION IS NOT ALL-TRAIN, AND THIS IS THE FIRST ENTRY ON THIS SCOPE WHERE IT IS NOT. Six train: `api-endpoint` 0x3b = 59, `blade-ui` 0x78 = 120, `flux` 0xa2 = 162, `livewire` 0xd9 = 217, `react-shadcn-ui` 0x47 = 71, `sql-writing` 0xf0 = 240, all above the holdout ceiling of 51. One HOLDOUT: `dashboard-design` is 0x04 = 4, far below it, so the partition rule sealed it and nobody chose that -- the property the name-hash rule exists to provide. Every entry from 2026-10-07 onward could write `the holdout is unchanged`; this one cannot. The sealed set moves 27 -> 28 and the realised split is 28/134 = 20.9 % against the rule's 19.9 % target. The holdout pin IS touched, and for a reason none of the recent entries had: a sealed ROW IS ADDED rather than a count line moved. All 27 previously sealed rows and all 100 previously published train rows reproduce BYTE-FOR-BYTE -- the diff against the tree reports exactly seven rows absent, ZERO hash mismatches and ZERO orphans, so no pre-existing corpus file was edited in this change and no holdout result taken against the earlier set is voided. The ordering claim stays scoped for the reason the 19th row established on 2026-10-01: a row authored today predates nothing, and a proposer evaluated against this holdout must either exclude `dashboard-design` or say it is using a row whose seal is younger than the evaluation. BOTH published train-side measurements are re-taken here on the terms the 2026-10-04, 2026-10-09, 2026-10-10, 2026-10-11 and 2026-10-12 entries set. routing-body-signal-verdict.json: verdict unchanged in kind, still `harmful` and still on the false-activation guard, but every input moved -- false activation 7.128 -> 5.917 pp against the unchanged +2.0 pp guard, delta-recall 4.517 -> 2.294 pp (still below the 5.0 pp bar, and now further below it than at any earlier re-take), McNemar p 0.07765 -> 0.3846, train corpora 100 -> 106, cases 964 -> 1030, catalogue unchanged at 300. delivery-set-measurement-2026-08-31.json: precision@5 81.016 -> 80.247 %, recall@5 62.218 -> 62.141 %, false activation 14.885 -> 15.779 %, adjudicated deliveries 374 -> 405, context cost 230.1 -> 230.6 tokens; the recall-loss ceiling stays BREACHED and the token target stays MET, so neither verdict changes in kind. Six train rows is why both moved; `dashboard-design` is sealed and contributes to neither. SEVEN entries were removed from trigger_eval_grandfather.json (173 -> 166). This is the same shape as the correction recorded against 2026-10-12 one entry above, and the difference is where it was caught: there the pair kept their entries, check_trigger_eval_presence exited 1 from the corpus commit onward and the trunk carried a red ratchet for a day; here the gate named all seven before the branch left the machine. The allowlist is shrink-only, so a skill that gains a corpus owes the removal in the same change -- and the easy way to miss it is the one that entry records: the file has four top-level keys and the allowlist is `skills`, so a reader taking the first array reads `_no_obligation` and gets False for every name. One date reads oddly on purpose: routing-body-signal-verdict.json stamps `measured_at` from UTC and the write landed at 23:xx UTC, so it says 2026-10-13 while this entry says 2026-10-14. The record is not stale; the two clocks disagree by two hours."
+        },
+        {
+            "date": "2026-10-14",
+            "scope": "skills",
+            "from": 0.4567,
+            "to": 0.4667,
+            "reason": "road-to-sentences-the-tree-can-back made three dated or unsourced claims defer to their source -- market-entry-analysis now requires a cited source on every market-size, growth-rate, price and competitor claim; positioning-strategy gained a Sourcing section over segment, category and alternative claims; and react-native-setup's Node floor reads the project's own .nvmrc / engines.node instead of a 2025 baseline that had already expired unnoticed. check_routing_coverage requires a corpus for every skill a diff touches and none of the three had one (137 -> 140 of 300). TEN cases each, the 2026-10-14b shape rather than the eleven of 2026-10-14: 5 exemplars of which one is German, 3 near-misses naming the neighbour each skill's own Do-NOT or cross-link list hands off to, and 2 counterexamples guarding that skill's own homonyms -- entry as a program entry point, market as an exchange feed, positioning as CSS, position as a camera placement, native as a database operator, expo as a trade show; 30 cases in all. The seed rises to the live value; it may never be LOWERED. ALL THREE ROWS SEAL, AND NO EARLIER ENTRY ON THIS SCOPE RECORDS THAT. market-entry-analysis is 0x00 = 0, positioning-strategy 0x04 = 4 and react-native-setup 0x21 = 33, every one below the holdout ceiling of 51. Three independent draws under the ceiling has probability (51/256)^3 ~ 0.79 %, so this is the unlikely arm of a fair rule and not a choice -- the directory names predate the corpus. 0x00 is the floor of the range and nothing can draw lower, which also falsifies a claim the 2026-10-14 entry makes in passing: 0x04 was not the lowest bucket ever drawn even then, because analysis-autonomous-mode and judge-spec-compliance sit at 0x02 and worktree-lifecycle at 0x03. The sealed set moves 28 -> 31, the train set does NOT move and stays at 109, and the realised split is 31/140 = 22.14 % against the rule's 19.9 % target, the widest it has stood. The holdout pin IS touched -- three sealed rows are ADDED -- and the ordering claim's scope moves with it: AC-6 still rests on the same 18 rows, so it is 18 of 31 rather than 18 of 28, and a proposer evaluated against this holdout must exclude the three or say it is using rows whose seal is younger than the evaluation. All 28 previously sealed rows and all 109 previously published train rows reproduce BYTE-FOR-BYTE: the diff against the tree reports exactly three rows absent, ZERO hash mismatches and ZERO orphans over the 137 rows published before this change, so no pre-existing corpus file was edited and no holdout result taken against the earlier set is voided. SET-SHA256 8446fbd4... -> bb76aad9..., recomputed by the artefact's own shell recipe rather than read off an assertion, and cross-checked against an independent implementation that agreed. BOTH published train-side measurements were re-taken although a holdout-only growth does not owe them (Growth 2026-10-01 says so explicitly). They were run because 'the partition says they cannot have moved' is a prediction, and a published pin is measured rather than predicted: measure_routing_signal --write and measure_delivery_sets --write each rewrote their file and diff against the pre-run copy reports ZERO bytes changed. routing-body-signal-verdict.json is unchanged in kind and in every input -- still harmful, false activation +5.94 pp against the +2.0 pp guard, delta-recall +1.86 pp against the 5.0 pp bar, McNemar p 0.4823, train corpora 109, cases 1060, catalogue 300. delivery-set-measurement-2026-08-31.json likewise: precision@5 80.10 %, recall@5 62.08 %, false activation@5 15.90 %. Three grandfather entries removed with the corpora: 163 -> 160, the shrink-only direction, line-scoped so the file's escaped em-dashes stay as they were. One pre-existing gap closed in passing: the artefact's Train table held 108 rows against 109 train corpora because `database` was published only as a malformed cell in the Correction 2026-08-31 table and never reached the Train table; its hash there matches the tree, so nothing was unpinned and no hash moves. ONE GAP IS LEFT OPEN AND NAMED: the 0.4467 -> 0.4567 step earlier today moved seed and seed_provenance without writing a history row, so the chain below reads 0.4233 -> 0.4467 and then 0.4567 -> 0.4667. The missing row is that change's to record, not this one's, and inventing it here would publish a reason this session did not take."
+        },
+        {
+            "date": "2026-10-14",
+            "scope": "skills",
+            "from": 0.4667,
+            "to": 0.47,
+            "reason": "road-to-account-lifecycle-threat-rows Phase 3.1 gave `security` the `triggers:` block and the trigger corpus it had never carried (140 -> 141 of 300). Unlike the 2026-10-12 entry above, this corpus is the STEP's own deliverable rather than a gate's remedy: the step exists because 'add login with Google' was a should-NOT for authz-review and a should-trigger for no skill in the tree, so the account-lifecycle rows this roadmap adds to the threat corpus had no skill a consumer would reach them through. Sixteen cases -- nine exemplars covering sign-up, verification, reset, password storage, MFA, passkey, session rotation, logout and the Google-login query itself, five near-misses that are this skill's own four written 'Do NOT use when' routes plus request validation, and two counterexamples carrying no security vocabulary at all. The seed rises to the live value; it may never be LOWERED. The partition is TRAIN, not holdout: sha256('security')[0:2] is 0x5d = 93, above the holdout ceiling of 51, so BOTH published train-side measurements are re-taken here on the terms the 2026-10-04 through 2026-10-14c entries set. THE FROM-VALUE AND EVERY FIGURE IN THIS ENTRY WERE RE-TAKEN ON 2026-10-14 AFTER MERGING origin/main, AND THE ORIGINAL TAKE IS NOT PRESERVED BESIDE THEM, because it was never published: the branch measured 127 -> 128 of 300 against a trunk that had since reached 140, and a from-value chaining off a superseded trunk is a wrong number rather than an instructive one. What the original take DID record and is kept: the measurement was taken twice on the branch because the first ran after the corpus landed but before 3.2 added fifteen lines to the same skill's BODY, and the body arm reads the body -- a train-side measurement is valid only when every edit to a train skill, corpus and body alike, is already on disk. The holdout pin IS touched for the row-count reason only -- one train row added, no holdout row's content changed -- so the sealed 31 rows are byte-identical and the ordering claim is untouched. ONE entry was removed from trigger_eval_grandfather.json with this corpus. The branch's own earlier note claimed three removals (`security`, `form-handler`, `judge-security-auditor`); the latter two were already removed on the trunk by drain/local-gates-wired-20261013, which the 2026-10-12 entry's correction above records, so only `security` was left to remove and the merge kept exactly that one."
+        },
+        {
+            "date": "2026-10-14",
+            "scope": "skills",
+            "from": 0.47,
+            "to": 0.4733,
+            "reason": "road-to-account-surface-detector step 3.2 wired check_account_surface into launch-readiness AUTH area (and into the production-validator subagent, which carries no corpus), so check_routing_coverage required a corpus from the touched skill and it had none (141 -> 142 of 300). TEN cases, the 2026-10-14b shape: 5 exemplars of which one is German covering the go/no-go decision, the pre-merge checklist, the rollout plan with rollback criteria, the prelaunch diagnostic and the ops handoff; 3 near-misses that are the skill own WHEN-NOT routes (release-comms, finishing-a-development-branch, stakeholder-tradeoff); 2 counterexamples guarding launch and rollback outside a release. The seed rises to the live value; it may never be LOWERED. The partition is TRAIN: sha256(launch-readiness)[0:2] is 0x71 = 113, above the holdout ceiling of 51, so the sealed 31 rows are untouched and BOTH published train-side measurements are re-taken in the same change, after the skill body edit was on disk. One grandfather entry removed with the corpus: 159 -> 158 as check_trigger_eval_presence reads it, the shrink-only direction."
+        },
+        {
+            "date": "2026-10-16",
+            "scope": "skills",
+            "from": 0.4733,
+            "to": 0.4933,
+            "reason": "road-to-skills-that-arrive-when-called step 1.1 renamed the package frontmatter key context: large to context_window: large on the six skills that carried it (deep-reading-analyst, project-analysis-core, project-analysis-hypothesis-driven, project-analyzer, repomix-packer, universal-project-analysis), so the host-owned context key is free. check_routing_coverage requires a corpus from every touched skill and none of the six had one (142 -> 148 of 300). SIX cases each, smaller than the 2026-10-14b shape and stated as such: 3 exemplars of which one is German, 2 near-misses naming the neighbour each skill's own Do-NOT list hands off to, 1 counterexample. The seed rises to the live value; it may never be LOWERED. ONE ROW SEALS: deep-reading-analyst is 0x19 = 25, below the holdout ceiling of 51; the other five train. The sealed set moves 31 -> 32 and the train set 111 -> 116. BOTH published train-side measurements were re-taken after every skill edit was on disk: routing-body-signal-verdict.json still harmful (false activation +5.829 pp against the +2.0 pp guard, delta-recall +1.94 pp against the 5.0 pp bar, McNemar p 0.4445, train corpora 116, cases 1116, catalogue 300); delivery-set-measurement-2026-08-31.json precision@5 80.6 %, recall@5 61.552 %, false activation@5 15.301 %. Six grandfather entries removed with the corpora: 158 -> 152, the shrink-only direction."
+        },
+        {
+            "date": "2026-10-16",
+            "scope": "skills",
+            "from": 0.4933,
+            "to": 0.5033,
+            "reason": "road-to-steps-the-skill-cannot-skip steps 1.2 and 2.1 repaired reference links in design-review and roadmap-management and aligned skill-reviewer's split rule, so check_routing_coverage required a corpus from each touched skill and none of the three had one (148 -> 151 of 300). TEN cases each, the 2026-10-14b shape: 5 exemplars of which one is German, 3 near-misses, 2 counterexamples. The seed rises to the live value; it may never be LOWERED. TWO ROWS SEAL: design-review is 0x1d = 29 and roadmap-management 0x07 = 7, below the holdout ceiling of 51; skill-reviewer is 0x51 = 81 and trains. The sealed set moves 32 -> 34 and the train set 116 -> 117. BOTH published train-side measurements were re-taken after every skill edit was on disk: routing-body-signal-verdict.json still harmful (false activation +5.776 pp against the +2.0 pp guard, delta-recall +2.269 pp against the 5.0 pp bar, McNemar p 0.3616, train corpora 117, cases 1127, catalogue 300); delivery-set-measurement-2026-08-31.json precision@5 80.645 %, recall@5 61.082 %, false activation@5 15.162 %. Three grandfather entries removed with the corpora: 152 -> 149, the shrink-only direction."
+        },
+        {
+            "date": "2026-10-18",
+            "scope": "skills",
+            "from": 0.5033,
+            "to": 0.52,
+            "reason": "RECONSTRUCTED 2026-10-22 by the 17.0.0 self-review, because the change that moved the seed added no row and the chain jumped 0.5033 -> 0.52 unrecorded. git log -S'\"skills\": 0.52' names 3a9b54d4c (#2449, brand checkability), which authored corpus files for content-funnel-design, gtm-launch, messaging-architecture, release-comms and voice-and-tone-design (151 -> 156 of 300). Recorded from the commit, not re-measured on that tree."
+        },
+        {
+            "date": "2026-10-18",
+            "scope": "skills",
+            "from": 0.52,
+            "to": 0.5216,
+            "reason": "road-to-type-with-a-witness Phase 4 added the `type-design` skill, which moved the DENOMINATOR 300 -> 301 -- the ratio-falls-two-ways case the _comment names. The repair is the corpus, not the seed: the skill shipped its own evals/triggers.json in the same change (156 -> 157 of 301), so the ratio ROSE across the increase. 12 cases: 6 exemplars of which one is German, 5 near-misses naming the neighbours and 1 counterexample. sha256('type-design')[0:2] is 0x43 = 67, above the holdout ceiling of 51, so it trains (121 -> 122) and the sealed set stays at 35. BOTH published train-side measurements were re-taken after every skill edit was on disk: routing-body-signal-verdict.json still harmful (false activation +5.73 pp against the +2.0 pp guard, delta-recall +3.67 pp against the 5.0 pp bar, McNemar p 0.1214, train corpora 122, cases 1175, catalogue 301); delivery-set-measurement-2026-08-31.json precision@5 80.81 %, recall@5 59.77 %, false activation@5 14.76 %. The seed rises to the live value; it may never be LOWERED."
+        },
+        {
+            "date": "2026-10-18",
+            "scope": "skills",
+            "from": 0.5216,
+            "to": 0.5548,
+            "reason": "road-to-steps-the-skill-cannot-skip AC-1 repaired dead SKILL.md links in ten skills that carried no corpus, so check_routing_coverage required one from each (157 -> 167 of 301). TEN cases each: 5 exemplars of which one is German, 3 near-misses, 2 counterexamples. ONE ROW SEALS: condense-memory is 0x01 = 1, below the holdout ceiling of 51; the other nine train. The sealed set moves 35 -> 36 and the train set 122 -> 131. BOTH published train-side measurements were re-taken after every skill edit was on disk: routing-body-signal-verdict.json still harmful (false activation +5.636 pp against the +2.0 pp guard, delta-recall +3.106 pp against the 5.0 pp bar, McNemar p 0.1768, train corpora 131, cases 1265, catalogue 301); delivery-set-measurement-2026-08-31.json precision@5 80.793 %, recall@5 60.093 %, false activation@5 14.815 %. Ten grandfather entries removed with the corpora, the shrink-only direction. The seed rises to the live value; it may never be LOWERED."
+        },
+        {
+            "date": "2026-10-19",
+            "scope": "skills",
+            "from": 0.5548,
+            "to": 0.5581,
+            "reason": "road-to-a-canvas-the-code-owns step 7.3 added one sentence to playwright-testing's SKILL.md, so check_routing_coverage required a corpus from the touched skill, which had none (167 -> 168 of 301). TEN cases: 5 exemplars of which one is German, 3 near-misses (playwright-architect, mobile-e2e-strategy, api-testing), 2 counterexamples. sha256('playwright-testing')[0:2] is 0x61 = 97, above the holdout ceiling of 51, so it trains (131 -> 132) and the sealed set stays at 36. BOTH published train-side measurements were re-taken after the skill edit was on disk: routing-body-signal-verdict.json still harmful (false activation +5.751 pp against the +2.0 pp guard, delta-recall +2.928 pp against the 5.0 pp bar, McNemar p 0.2018, train corpora 132, cases 1275, catalogue 301); delivery-set-measurement-2026-08-31.json precision@5 80.952 %, recall@5 60.247 %, false activation@5 14.696 %. One grandfather entry removed with the corpus, the shrink-only direction. The seed rises to the live value; it may never be LOWERED."
+        },
+        {
+            "date": "2026-10-20",
+            "scope": "skills",
+            "from": 0.5581,
+            "to": 0.5615,
+            "reason": "road-to-an-analysis-family-that-deepens-itself step 6.2b renamed the source door analyze:repo -> analyze:reference, which rewrote two routing-away links in competitive-positioning's SKILL.md, so check_routing_coverage required a corpus from the touched skill, which had none (168 -> 169 of 301). SEVEN cases: 3 exemplars of which one is German, 3 near-misses taken from that skill's own When-to-use routing-away list (analyze-reference, upstream-contribute, dcf-modeling), 1 counterexample. sha256('competitive-positioning')[0:2] is 0x21 = 33, BELOW the holdout ceiling of 51, so the partition rule SEALS it: the sealed set moves 36 -> 37 and the train set stays at 132. The two published train-side measurements were therefore NOT re-taken, and that is the partition rather than a skipped step -- both are computed over the TRAIN corpora, this growth adds none and edits none, so routing-body-signal-verdict.json (still harmful; train corpora 132, cases 1275, catalogue 301) and delivery-set-measurement-2026-08-31.json (precision@5 80.952 %, recall@5 60.247 %, false activation@5 14.696 %) carry over byte-identical inputs and unchanged figures. The seed rises to the live value; it may never be LOWERED."
+        },
+        {
+            "date": "2026-10-21",
+            "scope": "skills",
+            "from": 0.5615,
+            "to": 0.5714,
+            "reason": "road-to-security-fixes-carried-to-the-merge and road-to-advisories-heard-the-same-day edited composer-packages (a Composer audit section), dependency-upgrade (the security-fix procedure) and docker (an example image pin), so check_routing_coverage required a corpus from each touched skill (169 -> 172 of 301). Ten cases each: 5 exemplars of which one is German, 3 near-misses, 2 counterexamples. All three TRAIN (0x53, 0x5a, 0xd5, above the ceiling of 51): the sealed set stays at 37, the train set moves 132 -> 135, and both train-side measurements were re-taken in the same change (routing-body-signal-verdict.json still harmful, +5.77 pp false activation; delivery-set-measurement-2026-08-31.json precision@5 80.85 %). The seed rises to the live value; it may never be LOWERED."
         }
     ]
 }