@event4u/agent-config 14.12.0 → 14.14.0

package/src/config/rule-stub-ceilings.json

@@ -1,115 +1,136 @@
 {
-  "_comment": "Per-rule stub ceilings (road-to-rule-stub-projection Phase 2). One exact-BPE token ceiling per rule carrying a migration pointer. SHRINK-ONLY: a rule may fall below its ceiling freely, and may not rise above it. Raising a ceiling to clear a failing check is the config-weakening move this repo blocks by construction \u2014 every raise needs a `history` entry whose `reason` is a real sentence. The ceilings are a floor-plus-residue SNAPSHOT taken at adoption, not a target: a rule at its ceiling is not thereby correct, only not worse. Regenerate with `check_rule_stub_ceiling --write-baseline`; never hand-edit the counts.",
-  "measured_at_commit": "178b89235",
-  "tokens_exact": true,
-  "ceilings": {
-    "active-remediation": 1601,
-    "analysis-skill-routing": 76,
-    "architecture": 955,
-    "artifact-drafting-protocol": 680,
-    "augment-edit-discipline": 137,
-    "autonomous-execution": 1633,
-    "brand-consistency": 133,
-    "broken-access-control": 743,
-    "cli-output-handling": 83,
-    "code-comment-discipline": 692,
-    "commit-conventions": 136,
-    "context-hygiene": 2470,
-    "copilot-routing": 128,
-    "design-fidelity": 2540,
-    "devcontainer-routing": 128,
-    "docker-commands": 79,
-    "domain-adoption-policy": 497,
-    "framework-neutrality-in-generic-skills": 636,
-    "git-history-discipline": 1205,
-    "improve-before-implement": 845,
-    "laravel-routing": 126,
-    "laravel-translations": 82,
-    "legal-safety-floor": 1401,
-    "linked-projects-onboarding-gate": 172,
-    "minimal-safe-diff": 893,
-    "missing-tool-handling": 103,
-    "model-recommendation": 156,
-    "no-roadmap-references": 644,
-    "package-ci-checks": 74,
-    "persona-governance": 330,
-    "php-coding": 80,
-    "provider-lifecycle-discipline": 472,
-    "reviewer-awareness": 131,
-    "roadmap-ci-steps-policy": 564,
-    "roadmap-progress-sync": 2479,
-    "rule-type-governance": 87,
-    "skill-improvement-trigger": 81,
-    "skill-quality": 78,
-    "slash-command-routing-policy": 77,
-    "symfony-routing": 129,
-    "ui-audit-gate": 1491,
-    "untrusted-input-defense": 934,
-    "upstream-proposal": 76
-  },
-  "history": [
-    {
-      "date": "2026-08-19",
-      "rule": "roadmap-progress-sync",
-      "from": 1424,
-      "to": 2117,
-      "reason": "Iron Law 3 gained the preservation test that decides WHO resolves a deferred item \u2014 a routing table plus the two council refinements that make it binding (a follow-up naming a roadmap that does not exist yet fails closed to the user; a council resolution is recorded at the item or it did not happen) and the residual limit both authoring seats named. This is obligation surface, not explanation: the table decides a route and the two refinements are the conditions under which the route is legal, so a reader who reaches only the pointer target has already made the decision. It could not move to the pointer target either \u2014 roadmap-progress-mechanics.md sits at the 16,000-char check_depth_budget ceiling in this same change, and pushing prose there reds a count ratchet whose whole job is refusing a fifth over-ceiling file. The section was condensed once before this raise (2201 -> 2117) by turning four prose paragraphs into bullets; further cutting would delete substance, which preservation-guard forbids."
+    "_comment": "Per-rule stub ceilings (road-to-rule-stub-projection Phase 2). One exact-BPE token ceiling per rule carrying a migration pointer. SHRINK-ONLY: a rule may fall below its ceiling freely, and may not rise above it. Raising a ceiling to clear a failing check is the config-weakening move this repo blocks by construction \u2014 every raise needs a `history` entry whose `reason` is a real sentence. The ceilings are a floor-plus-residue SNAPSHOT taken at adoption, not a target: a rule at its ceiling is not thereby correct, only not worse. Regenerate with `check_rule_stub_ceiling --write-baseline`; never hand-edit the counts.",
+    "measured_at_commit": "580cb1138",
+    "tokens_exact": true,
+    "ceilings": {
+        "active-remediation": 1595,
+        "analysis-skill-routing": 76,
+        "architecture": 955,
+        "artifact-drafting-protocol": 690,
+        "augment-edit-discipline": 137,
+        "autonomous-execution": 1633,
+        "brand-consistency": 133,
+        "broken-access-control": 743,
+        "cli-output-handling": 83,
+        "code-comment-discipline": 750,
+        "commit-conventions": 136,
+        "context-hygiene": 2470,
+        "copilot-routing": 128,
+        "design-fidelity": 2540,
+        "devcontainer-routing": 128,
+        "docker-commands": 79,
+        "domain-adoption-policy": 497,
+        "framework-neutrality-in-generic-skills": 636,
+        "git-history-discipline": 1205,
+        "improve-before-implement": 1385,
+        "laravel-routing": 126,
+        "laravel-translations": 82,
+        "legal-safety-floor": 1401,
+        "linked-projects-onboarding-gate": 172,
+        "minimal-safe-diff": 888,
+        "missing-tool-handling": 103,
+        "model-recommendation": 156,
+        "no-roadmap-references": 644,
+        "package-ci-checks": 74,
+        "persona-governance": 330,
+        "php-coding": 80,
+        "provider-lifecycle-discipline": 472,
+        "reviewer-awareness": 131,
+        "roadmap-ci-steps-policy": 564,
+        "roadmap-progress-sync": 2479,
+        "rule-type-governance": 87,
+        "skill-improvement-trigger": 81,
+        "skill-quality": 78,
+        "slash-command-routing-policy": 77,
+        "symfony-routing": 129,
+        "ui-audit-gate": 1491,
+        "untrusted-input-defense": 934,
+        "upstream-proposal": 76
     },
-    {
-      "date": "2026-08-20",
-      "rule": "context-hygiene",
-      "from": 2447,
-      "to": 2470,
-      "reason": "One sentence, +23 tokens, and it is obligation surface rather than explanation: for CI, do not hand-write the wait loop. The mechanics went to the pointer target where this rule's own migration note already sends them (context-hygiene-mechanics gains the exit codes, the four not-settled states and the measured incident) \u2014 that move is what took the growth from +179 to +23, and the residue is the pointer itself, which cannot live in the guideline because a rule that does not name the helper is a rule nobody follows to it. The incident it encodes: a hand-written `until ! gh pr checks N | grep -q pending` exited when the API failed, because the error text contains no `pending`, and the session reported a settle that had not happened and retracted it in the next reply. The rule's neighbouring waiter law already says one waiter per condition; what it did not say is that the exit condition is the part that gets written wrong, which is why the helper is named here rather than described."
-    },
-    {
-      "date": "2026-08-20",
-      "rule": "roadmap-progress-sync",
-      "from": 2117,
-      "to": 2185,
-      "reason": "One sentence in Iron Law 1, +68 tokens, and it is obligation surface rather than explanation: `roadmap:progress` now passes `--archive`, so the law's operational meaning changed \u2014 a complete roadmap is moved by the same run instead of reported to a human, and the two exclusions a reader must know before trusting that (`--check` never archives, the PostToolUse hook never archives) are exactly the part that cannot be deferred to a pointer. The prose was NOT moved to docs/guidelines/agent-infra/roadmap-progress-mechanics.md, which is where this rule's migration note sends its mechanics, because that file has roughly a hundred characters of headroom under the 16,000-char depth ceiling and check_depth_budget is shrink-only with four files already over it \u2014 a FIFTH over-ceiling file reds, so moving 400 characters there would have traded this gate for that one. The full contract, the council record and its recorded dissent live with the code in src/agent-src/scripts/archival_sweep.ts, which is the third home and the one with room."
-    },
-    {
-      "date": "2026-08-20",
-      "rule": "artifact-drafting-protocol",
-      "from": 611,
-      "to": 680,
-      "reason": "Three lines plus a pointer extension, +69 tokens, for a new contract field that governs what a run may author without asking: an artifact discovered AFTER an accepted execution contract halts by default, and `late_artifacts: auto-research` is the only declared value that continues, capped at three per run. The five-step auto-research procedure, the reasoning for the default and its reversibility argument all went to the pointer target (artifact-drafting-protocol-mechanics gains a Late artifacts section) rather than here, which is what held the growth to a bullet. The residue is obligation surface and cannot live in the guideline: this rule's own batch-mode bullet already tells the reader that a mid-run artifact triggers the full interactive protocol, so a reader who never opens the guideline would conclude that a declared value cannot change it. Named default rather than described behaviour, because the default direction is the decision -- AI council 2/2 (anthropic + openai) chose `halt` over `auto-research` on the autonomy-defaults-sheet fork of road-to-user-out-of-the-loop, record agents/evidence/council/drain-blocker-dispositions-a.md."
-    },
-    {
-      "date": "2026-08-20",
-      "rule": "roadmap-progress-sync",
-      "from": 2185,
-      "to": 2479,
-      "reason": "One subsection, +294 tokens, for the `deferred_policy` contract field: a declared policy removes the synchronous options ROUND at closure without moving any row of the preservation test's routing table. It is obligation surface and it cannot live anywhere else. Not in the pointer target -- roadmap-progress-mechanics.md sits at 16,039 chars against check_depth_budget's 16,000-char ceiling, and that gate is a count ratchet whose next failure is a FIFTH over-ceiling file, so prose pushed there is the one move that turns a warn into a red. Not in the contract context either: Iron Law 3 is the gate being amended, and a reader who learns from roadmap-execution-contract that a field exists but never reads this rule would not know that `cancel-with-memo` still stops for the user. The three-row table is the delta in full; the reasoning, both alternatives and the reversibility argument went to the contract context's own decision comment. Condensed once before this raise (2509 -> 2411) by cutting a paragraph that restated the preservation test. AI council 2/2 (anthropic + openai) on the autonomy-defaults-sheet fork of road-to-user-out-of-the-loop, record agents/evidence/council/drain-blocker-dispositions-a.md. RE-BASED on merge with origin/main, 2026-08-20: this entry was authored against a base of 2117 and raised to 2411. main had independently raised the same ceiling 2117 -> 2185 for its own reason, so the merged tree measures 2479 and the honest arithmetic for THIS raise is 2185 -> 2479 -- the same +294 for the same subsection, re-based rather than re-argued. Recorded rather than silently regenerated because the gate's own failure text names a history-versus-ceilings disagreement as the signature of a hand-edited count, and a merge that quietly adopted one side's numbers would produce exactly that signature with nothing saying why."
-    },
-    {
-      "date": "2026-08-22",
-      "rule": "untrusted-input-defense",
-      "from": 845,
-      "to": 934,
-      "reason": "Six lines, +89 tokens, correcting a claim the section itself had made false. It said no content-scanning backstop existed and that a future one would be the first candidate to change enforced_by; injection_scan_hook.ts now exists, so the first half is false, and the second was always wrong because a warn-only hook cannot refuse and therefore cannot enforce. Every sentence left carries one of those claims. The measurement half -- slot, 15 channels, 99.00 percent recall, 0.85 percent false positives, the manifest keys -- did move to the pointer target, which is what cut the overshoot from +194 to +89; what stays is the rule's own enforcement disclosure, and a stub ceiling that forbids correcting a false statement about the rule's own enforcement is the wrong constraint applied to the right file."
-    },
-    {
-      "date": "2026-08-23",
-      "rule": "design-fidelity",
-      "from": 2245,
-      "to": 2540,
-      "reason": "One subsection, +295 tokens, for the second axis: artefact MATURITY is a property of the handed-over artefact, instruction MANDATE is a property of the prose, and `strict` on a self-declared wireframe is a structure mandate rather than a pixel mandate. It is an Iron Law plus the two discriminator clauses that decide it -- obligation surface, which this repository's rule/mechanics split puts in the rule by construction. The lookup half ALREADY went to the pointer target in the same change: the maturity-to-spec table and the pinned `daf-wireframe-near-miss` row live in design-fidelity-mechanics.md \u00a7 Artefact maturity, and the rule links them. Pushing the remaining prose there too is the one move that is not available -- that migration is exactly what took the mechanics file from 15,265 to 16,524 chars past check_depth_budget's 16,000 ceiling, and paying for it required splitting the asset & imagery block out into its own guideline in this same commit series. A second push would re-red the ratchet the split just cleared. Raised rather than condensed because the two clauses that would compress are the ones that carry the decision: 'the discriminator reads the ARTEFACT, never the prose' and 'an undeclared artefact is treated as finished' -- drop either and the rule authorises the redesign it exists to prevent."
-    },
-    {
-      "date": "2026-08-23",
-      "rule": "ui-audit-gate",
-      "from": 1280,
-      "to": 1491,
-      "reason": "road-to-frontend-power E2.2 moved this rule's `enforced_by:` from `none` to `hook:design-pass`, because `agent-config ui:audit` now writes the artefact the rule previously had nothing to point at. +127 tokens, and every one of them is the SCOPE of that change rather than a description of the mechanism: the check is FRESHNESS not existence (a stale artefact looks like evidence, which is worse than a missing one), the carrier is default-OFF and warns before it blocks, and the greenfield asymmetry is NOT closed -- `_lib/ui_surface.ts` is a path predicate, measured recall 20/23 with all three misses being a UI request that has produced no UI file yet, so this gate cannot fire on the first write of a new surface. Condensed first, from +279 to +127, by cutting the narrative framing and the changed-in-both-directions paragraph. The three limits are what remains and they are the reason the raise is legitimate: a rule that newly claims a carrier and does not state where the carrier stops is a stronger claim than the tree supports, which is the exact over-claim this repository has already had to correct twice in the hook-architecture host table. Not pushed to the pointer target: `existing-ui-audit` is a SKILL describing how to run an audit, and the boundary of a gate's enforcement is an obligation-surface fact -- putting it there would file the limit where nobody reading the gate will find it. Amended in the same change, 1407 -> 1491 (+84): landing the carrier revealed that `design-pass-stop` is `severity: advisory` and the dispatcher enforces advisory as a CEILING, so an EXIT_BLOCK from it is downgraded and then mapped to exit 0. Claiming a block in this rule while the transport discards it would be the inert-refusal defect `run-continuation` already recorded once. Limit 2 therefore states what actually happens -- the stop verdict is REPORTED as would-block, not enforced -- and names what would make it real: an entry in concern_severity.test.ts's BLOCKING_ALLOWLIST, which that file calls a security-relevant decision and which would be the third turn-END refusal in the tree. That is a stronger reason to spend the tokens than the original raise had: a rule whose enforced_by names a carrier, and which does not say the carrier cannot refuse, is the over-claim this whole section exists to prevent."
-    },
-    {
-      "date": "2026-08-23",
-      "rule": "framework-neutrality-in-generic-skills",
-      "from": 621,
-      "to": 636,
-      "reason": "Two entries on the exempt list, +15 tokens: `^monorepo-workspace$` and `^workspace-link$`, the two skills road-to-monorepo-scope-and-detection adds. The list is not lookup material and cannot go to the pointer target: it is the rule's own ACTIVATION condition -- the gate decides from it whether an artefact is generic at all, so a rule whose exempt list lives elsewhere cannot answer its own first question. Every prior addition to that list moved this number for the same reason, and the alternative the gate offers (move prose to the pointer target) has nothing to move here: the two entries ARE the change. Raised rather than condensed because 15 tokens of pattern cannot be condensed without deleting one of the two patterns, which would make a framework-specific skill fail a rule it is exempt from."
-    }
-  ]
+    "history": [
+        {
+            "date": "2026-08-19",
+            "rule": "roadmap-progress-sync",
+            "from": 1424,
+            "to": 2117,
+            "reason": "Iron Law 3 gained the preservation test that decides WHO resolves a deferred item \u2014 a routing table plus the two council refinements that make it binding (a follow-up naming a roadmap that does not exist yet fails closed to the user; a council resolution is recorded at the item or it did not happen) and the residual limit both authoring seats named. This is obligation surface, not explanation: the table decides a route and the two refinements are the conditions under which the route is legal, so a reader who reaches only the pointer target has already made the decision. It could not move to the pointer target either \u2014 roadmap-progress-mechanics.md sits at the 16,000-char check_depth_budget ceiling in this same change, and pushing prose there reds a count ratchet whose whole job is refusing a fifth over-ceiling file. The section was condensed once before this raise (2201 -> 2117) by turning four prose paragraphs into bullets; further cutting would delete substance, which preservation-guard forbids."
+        },
+        {
+            "date": "2026-08-20",
+            "rule": "context-hygiene",
+            "from": 2447,
+            "to": 2470,
+            "reason": "One sentence, +23 tokens, and it is obligation surface rather than explanation: for CI, do not hand-write the wait loop. The mechanics went to the pointer target where this rule's own migration note already sends them (context-hygiene-mechanics gains the exit codes, the four not-settled states and the measured incident) \u2014 that move is what took the growth from +179 to +23, and the residue is the pointer itself, which cannot live in the guideline because a rule that does not name the helper is a rule nobody follows to it. The incident it encodes: a hand-written `until ! gh pr checks N | grep -q pending` exited when the API failed, because the error text contains no `pending`, and the session reported a settle that had not happened and retracted it in the next reply. The rule's neighbouring waiter law already says one waiter per condition; what it did not say is that the exit condition is the part that gets written wrong, which is why the helper is named here rather than described."
+        },
+        {
+            "date": "2026-08-20",
+            "rule": "roadmap-progress-sync",
+            "from": 2117,
+            "to": 2185,
+            "reason": "One sentence in Iron Law 1, +68 tokens, and it is obligation surface rather than explanation: `roadmap:progress` now passes `--archive`, so the law's operational meaning changed \u2014 a complete roadmap is moved by the same run instead of reported to a human, and the two exclusions a reader must know before trusting that (`--check` never archives, the PostToolUse hook never archives) are exactly the part that cannot be deferred to a pointer. The prose was NOT moved to docs/guidelines/agent-infra/roadmap-progress-mechanics.md, which is where this rule's migration note sends its mechanics, because that file has roughly a hundred characters of headroom under the 16,000-char depth ceiling and check_depth_budget is shrink-only with four files already over it \u2014 a FIFTH over-ceiling file reds, so moving 400 characters there would have traded this gate for that one. The full contract, the council record and its recorded dissent live with the code in src/agent-src/scripts/archival_sweep.ts, which is the third home and the one with room."
+        },
+        {
+            "date": "2026-08-20",
+            "rule": "artifact-drafting-protocol",
+            "from": 611,
+            "to": 680,
+            "reason": "Three lines plus a pointer extension, +69 tokens, for a new contract field that governs what a run may author without asking: an artifact discovered AFTER an accepted execution contract halts by default, and `late_artifacts: auto-research` is the only declared value that continues, capped at three per run. The five-step auto-research procedure, the reasoning for the default and its reversibility argument all went to the pointer target (artifact-drafting-protocol-mechanics gains a Late artifacts section) rather than here, which is what held the growth to a bullet. The residue is obligation surface and cannot live in the guideline: this rule's own batch-mode bullet already tells the reader that a mid-run artifact triggers the full interactive protocol, so a reader who never opens the guideline would conclude that a declared value cannot change it. Named default rather than described behaviour, because the default direction is the decision -- AI council 2/2 (anthropic + openai) chose `halt` over `auto-research` on the autonomy-defaults-sheet fork of road-to-user-out-of-the-loop, record agents/evidence/council/drain-blocker-dispositions-a.md."
+        },
+        {
+            "date": "2026-08-20",
+            "rule": "roadmap-progress-sync",
+            "from": 2185,
+            "to": 2479,
+            "reason": "One subsection, +294 tokens, for the `deferred_policy` contract field: a declared policy removes the synchronous options ROUND at closure without moving any row of the preservation test's routing table. It is obligation surface and it cannot live anywhere else. Not in the pointer target -- roadmap-progress-mechanics.md sits at 16,039 chars against check_depth_budget's 16,000-char ceiling, and that gate is a count ratchet whose next failure is a FIFTH over-ceiling file, so prose pushed there is the one move that turns a warn into a red. Not in the contract context either: Iron Law 3 is the gate being amended, and a reader who learns from roadmap-execution-contract that a field exists but never reads this rule would not know that `cancel-with-memo` still stops for the user. The three-row table is the delta in full; the reasoning, both alternatives and the reversibility argument went to the contract context's own decision comment. Condensed once before this raise (2509 -> 2411) by cutting a paragraph that restated the preservation test. AI council 2/2 (anthropic + openai) on the autonomy-defaults-sheet fork of road-to-user-out-of-the-loop, record agents/evidence/council/drain-blocker-dispositions-a.md. RE-BASED on merge with origin/main, 2026-08-20: this entry was authored against a base of 2117 and raised to 2411. main had independently raised the same ceiling 2117 -> 2185 for its own reason, so the merged tree measures 2479 and the honest arithmetic for THIS raise is 2185 -> 2479 -- the same +294 for the same subsection, re-based rather than re-argued. Recorded rather than silently regenerated because the gate's own failure text names a history-versus-ceilings disagreement as the signature of a hand-edited count, and a merge that quietly adopted one side's numbers would produce exactly that signature with nothing saying why."
+        },
+        {
+            "date": "2026-08-22",
+            "rule": "untrusted-input-defense",
+            "from": 845,
+            "to": 934,
+            "reason": "Six lines, +89 tokens, correcting a claim the section itself had made false. It said no content-scanning backstop existed and that a future one would be the first candidate to change enforced_by; injection_scan_hook.ts now exists, so the first half is false, and the second was always wrong because a warn-only hook cannot refuse and therefore cannot enforce. Every sentence left carries one of those claims. The measurement half -- slot, 15 channels, 99.00 percent recall, 0.85 percent false positives, the manifest keys -- did move to the pointer target, which is what cut the overshoot from +194 to +89; what stays is the rule's own enforcement disclosure, and a stub ceiling that forbids correcting a false statement about the rule's own enforcement is the wrong constraint applied to the right file."
+        },
+        {
+            "date": "2026-08-23",
+            "rule": "design-fidelity",
+            "from": 2245,
+            "to": 2540,
+            "reason": "One subsection, +295 tokens, for the second axis: artifact MATURITY is a property of the handed-over artifact, instruction MANDATE is a property of the prose, and `strict` on a self-declared wireframe is a structure mandate rather than a pixel mandate. It is an Iron Law plus the two discriminator clauses that decide it -- obligation surface, which this repository's rule/mechanics split puts in the rule by construction. The lookup half ALREADY went to the pointer target in the same change: the maturity-to-spec table and the pinned `daf-wireframe-near-miss` row live in design-fidelity-mechanics.md \u00a7 Artifact maturity, and the rule links them. Pushing the remaining prose there too is the one move that is not available -- that migration is exactly what took the mechanics file from 15,265 to 16,524 chars past check_depth_budget's 16,000 ceiling, and paying for it required splitting the asset & imagery block out into its own guideline in this same commit series. A second push would re-red the ratchet the split just cleared. Raised rather than condensed because the two clauses that would compress are the ones that carry the decision: 'the discriminator reads the ARTEFACT, never the prose' and 'an undeclared artifact is treated as finished' -- drop either and the rule authorises the redesign it exists to prevent."
+        },
+        {
+            "date": "2026-08-23",
+            "rule": "ui-audit-gate",
+            "from": 1280,
+            "to": 1491,
+            "reason": "road-to-frontend-power E2.2 moved this rule's `enforced_by:` from `none` to `hook:design-pass`, because `agent-config ui:audit` now writes the artifact the rule previously had nothing to point at. +127 tokens, and every one of them is the SCOPE of that change rather than a description of the mechanism: the check is FRESHNESS not existence (a stale artifact looks like evidence, which is worse than a missing one), the carrier is default-OFF and warns before it blocks, and the greenfield asymmetry is NOT closed -- `_lib/ui_surface.ts` is a path predicate, measured recall 20/23 with all three misses being a UI request that has produced no UI file yet, so this gate cannot fire on the first write of a new surface. Condensed first, from +279 to +127, by cutting the narrative framing and the changed-in-both-directions paragraph. The three limits are what remains and they are the reason the raise is legitimate: a rule that newly claims a carrier and does not state where the carrier stops is a stronger claim than the tree supports, which is the exact over-claim this repository has already had to correct twice in the hook-architecture host table. Not pushed to the pointer target: `existing-ui-audit` is a SKILL describing how to run an audit, and the boundary of a gate's enforcement is an obligation-surface fact -- putting it there would file the limit where nobody reading the gate will find it. Amended in the same change, 1407 -> 1491 (+84): landing the carrier revealed that `design-pass-stop` is `severity: advisory` and the dispatcher enforces advisory as a CEILING, so an EXIT_BLOCK from it is downgraded and then mapped to exit 0. Claiming a block in this rule while the transport discards it would be the inert-refusal defect `run-continuation` already recorded once. Limit 2 therefore states what actually happens -- the stop verdict is REPORTED as would-block, not enforced -- and names what would make it real: an entry in concern_severity.test.ts's BLOCKING_ALLOWLIST, which that file calls a security-relevant decision and which would be the third turn-END refusal in the tree. That is a stronger reason to spend the tokens than the original raise had: a rule whose enforced_by names a carrier, and which does not say the carrier cannot refuse, is the over-claim this whole section exists to prevent."
+        },
+        {
+            "date": "2026-08-23",
+            "rule": "framework-neutrality-in-generic-skills",
+            "from": 621,
+            "to": 636,
+            "reason": "Two entries on the exempt list, +15 tokens: `^monorepo-workspace$` and `^workspace-link$`, the two skills road-to-monorepo-scope-and-detection adds. The list is not lookup material and cannot go to the pointer target: it is the rule's own ACTIVATION condition -- the gate decides from it whether an artifact is generic at all, so a rule whose exempt list lives elsewhere cannot answer its own first question. Every prior addition to that list moved this number for the same reason, and the alternative the gate offers (move prose to the pointer target) has nothing to move here: the two entries ARE the change. Raised rather than condensed because 15 tokens of pattern cannot be condensed without deleting one of the two patterns, which would make a framework-specific skill fail a rule it is exempt from."
+        },
+        {
+            "date": "2026-08-26",
+            "rule": "improve-before-implement",
+            "from": 845,
+            "to": 1385,
+            "reason": "road-to-evidence-gated-change Phase 3 lands D5 in this rule: the existence question was inert on bug fixes and trivial changes, which are the classes that most often re-add a helper the tree already has. +540 tokens, and each of the four additions is obligation surface rather than lookup material \u2014 the narrowed exclusion (the three heavy checks stay excluded, the one cheap question does not) with the diff-decidable skip that keeps it cheap, the six named verdicts with negative evidence owed for `new`, the two clauses that keep the set from degrading into a ranking or into similarity-driven abstraction, and the recorded reason the gate is not routed through /tdd. Migrating any of it to the pointer target is the one move that is not available: docs/guidelines/agent-infra/agent-interaction-and-decision-quality.md sits at 15,992 of the 16,000-char check_depth_budget ceiling \u2014 eight characters \u2014 and that gate is a shrink-only ratchet already carrying four baseline violations whose own note says a FIFTH over-ceiling file reds. Raised rather than condensed because the clauses that would compress are the ones carrying the decision: drop the heavy/cheap split and the question becomes a tax on the highest-frequency change class (risk-register row 3 of that roadmap), drop the negative-evidence clause and `new` becomes the default verdict again."
+        },
+        {
+            "date": "2026-08-27",
+            "rule": "artifact-drafting-protocol",
+            "from": 680,
+            "to": 690,
+            "reason": "road-to-composition-before-creation Phase 3.3 makes Phase B's verdict a WRITTEN record rather than only a reasoned one, so the rule has to name the field that carries it. +74 tokens, 680 -> 754, and the split was deliberate: everything that is lookup material \u2014 the field's YAML shape, the four dispositions with their discriminators, why `none_found` must exist so an author is never pushed to invent an incumbent, why `create_separate` is the value no incumbent vocabulary can express, and what the lint can and cannot check \u2014 went to the pointer target (`docs/guidelines/agent-infra/artifact-drafting-protocol-mechanics.md` \u00a7 Where the answer is written). What stayed is obligation surface and cannot be moved without the rule mandating a field it does not name: that the verdict is written into the artifact, the field's name, and the measurement behind it (1 of the 20 most recently added skills and rules carried a machine-readable record \u2014 `agents/evidence/analysis/authoring-search-record-sample-2026-08-27.md`). Net +10 (680 -> 690), and the drafting history is the part worth recording because the net hides it: the first draft added +248 (to 928) and FOUR shrink passes brought it back -- 928 -> 783 -> 754 -> 690 -- the last one forced by a different gate rather than chosen. An independent review read the earlier wording as contradicting the structured `from: 680, to: 690`; one seat called it cumbersome rather than wrong. Stated both ways now: initial draft +248, final net +10. `check_preamble_payload_budget` reds at 138,212 tokens -- its grace ceiling, set to the exact measured total on 2026-08-24 with zero slack and expiring 2026-11-10 -- and `origin/main` measures 138,195, so this branch had **17 tokens** of headroom for the whole rule payload, not 74. The addition was therefore folded into the existing Phase-B parenthetical instead of added as a paragraph, which is both cheaper and better placed: the field is now named where the phase is described. Measured after: 138,208, four under the ceiling. The ceiling here is the exact measured count with no slack, so the next growth is caught rather than absorbed."
+        },
+        {
+            "date": "2026-09-02",
+            "rule": "code-comment-discipline",
+            "from": 692,
+            "to": 750,
+            "reason": "Three Iron Law lines folded into the rule's existing block: a comment naming where the code came from is evidence in the wrong place, provenance is not a WHY, and a roadmap cannot license a comment this rule forbids. That is obligation surface \u2014 it decides what an author may write \u2014 and preservation-guard forbids condensing an Iron Law further. Everything explanatory went to the pointer target the rule already names: the measured case (509 German comment lines and 21 report- or provenance-shaped comments across 41 of 45 changed source files in a consumer repository, including a 101-line colour-analysis report inside a CSS file), the reason the rule's own keyword triggers miss the write, and the two enforcers. Three condensations took the growth from +546 to +58: move to the guideline, then fold into the existing Iron Law block rather than open a second section. The ceiling is set at the measured body rather than above it \u2014 headroom nobody has a use for is the spendable-surplus failure this file's own comment names."
+        }
+    ]
 }

package/src/config/turnaround-budget.json

@@ -0,0 +1,66 @@
+{
+    "_comment": "Turnaround baseline (road-to-agent-turnaround Phase 1.2/1.3). Produced by src/scripts/probe_turnaround.ts, NOT by the throwaway script that wrote agents/evidence/analysis/agent-turnaround-2026-08-30.md — see `instrument_vs_source` below, which records where the two disagree and why. EMPTY-CORPUS BEHAVIOUR, the whole of step 1.3: `empty_corpus: \"fail\"`, and the probe is deliberately NOT wired into `task ci` or any workflow. A transcript store is machine-local — `~/.claude/projects/<mangled-cwd>/*.jsonl` — so a fresh CI checkout has none, and a gate over it would be green-because-empty, which is the gate-that-scans-nothing-exits-green failure this repository names in src/config/gate-coverage.yml. Fail-closed plus local-only is the pair: the flag refuses an empty corpus rather than certifying it, and nothing schedules it where the corpus is always empty. Re-run it by hand, or from `task preflight` on a machine that has a store. THE MEASURING SESSION IS EXCLUDED, and that is correctness rather than hygiene: the first registered baseline (81.42) failed its own re-run minutes later at 81.61, because the probe was reading its own execution — API calls accumulating in the newest transcript against a user-request denominator of one. With the exclusion, two consecutive runs report 72.67 identically. `--include-current` opts back in for an offline corpus nobody is writing to.",
+    "schema_version": 1,
+    "registered_at": "2026-08-30",
+    "owner": "maintainer",
+    "review_by": "2027-02-28",
+    "instrument": "src/scripts/probe_turnaround.ts",
+    "corpus": {
+        "window": "the 10 most recently modified sessions in this package's transcript store, EXCLUDING the measuring session, read 2026-08-30",
+        "sessions": 10,
+        "user_requests": 42,
+        "api_calls": 3052,
+        "tool_calls": 3031,
+        "note": "The corpus is a MOVING WINDOW by construction — `--limit 10` over mtime — so a re-run is never the same ten sessions. That is why the baseline records the window's shape (sessions, requests, calls) alongside the ratios: a ratio that moves because the denominator changed is not a regression, and without the shape nobody can tell the two apart."
+    },
+    "baseline": {
+        "calls_per_request": 72.67,
+        "mean_batch_size": 1.01,
+        "blocking_share": 0.6202,
+        "context_floor_max": 230705
+    },
+    "directions": {
+        "calls_per_request": "REPORTED ONLY, not compared. Measured unstable twice for reasons that are not regressions: 81.42 -> 72.67 when the measuring session was excluded, then 72.67 -> 73.73 within the same afternoon because the corpus is an mtime window and concurrent sessions in the same project store slide into it. A ratchet on a number that moves when nothing changed teaches a reader to ignore the gate. The recorded value stays as the reference reading.",
+        "mean_batch_size": "UP — 1.00 is fully serial, so a HIGHER number is the improvement",
+        "blocking_share": "DOWN",
+        "context_floor_max": "DOWN"
+    },
+    "empty_corpus": "fail",
+    "instrument_vs_source": {
+        "context_floor": "AGREES. Instrument 217,385–230,705; source analysis 218,705–230,705. The upper bound is identical and the lower differs by 0.6 %, which is the strongest available evidence that the two readers parse the same store the same way.",
+        "blocking_share": "AGREES within the window difference: 60 % here, 64 % in the source.",
+        "mean_batch_size": "DISAGREES, and this is a FINDING rather than an instrument defect. The source reports 1.00 across 3,212 tool-using messages — 'not one message in the corpus carried two tool calls'. This window reports 1.01: 27 of 2,889 tool-using requests (0.93 %) carried 2–4 tool_use blocks, spread over five sessions. Parallel calls therefore DO occur, so 1.00 was a property of that window and not a floor with a mechanism behind it. Step 2.1 turns on exactly this.",
+        "calls_per_request": "DISAGREES LARGELY — 72.67 here against 42.6 in the source — and the difference is a DENOMINATOR effect, not a regression. This window contains long autonomous runs in which one user prompt drives hundreds of calls, so `user_requests` falls (42 vs 76) while `api_calls` stays comparable (3,052 vs 3,241). The metric is sensitive to session SHAPE, which is a real limitation of it as a ratchet and is recorded here rather than smoothed over: read it together with `corpus`, and treat a movement that tracks `user_requests` as no evidence at all."
+    },
+    "subsequent_readings": [
+        {
+            "read_at": "2026-08-30",
+            "reason": "road-to-turnaround-followups step 1.1 — re-read `mean_batch_size` against the 1.01 baseline after the batching obligation landed in `src/agent-src/contexts/communication/rules-auto/token-efficiency-mechanics.md` (commit af0cf0bf0, 2026-08-30 16:38:40 +0200 = 14:38:40Z).",
+            "invocation": "./scripts-run src/scripts/probe_turnaround --store ~/.claude/projects/-Users-mathiasberg-projects-galawork-galawork-packages-event4u-agent-config --limit 10 --against-baseline",
+            "invocation_note": "The roadmap's literal command carries no `--store`, and the probe defaults to `defaultStore(process.cwd())`. This reading was taken from the `drain/turnaround-followups` worktree at /private/tmp/ac-drain10, whose slug `-private-tmp-ac-drain10` names no directory — so the bare command measures ZERO sessions and exits 1 on `empty_corpus: \"fail\"`, correctly. `--store` was pointed at the store the FIRST reading measured, because a delta against a different corpus is not a delta.",
+            "corpus": {
+                "window": "the 10 most recently modified sessions in this package's transcript store, EXCLUDING the measuring session (dropped by CLAUDE_CODE_SESSION_ID 90ddc54d…, which the host exported), read 2026-08-30 ~20:20Z",
+                "sessions": 10,
+                "user_requests": 35,
+                "api_calls": 3281,
+                "tool_calls": 3266,
+                "tool_using_requests": 3237,
+                "note": "THE TEN-POST-CHANGE-SESSION PRECONDITION IS NOT MET, and this is the reading's principal finding rather than a caveat on it — risk 2 of the roadmap that ordered this step predicted exactly it. By first-transcript-timestamp, ONE of the ten sessions (4fb84fae, first 16:25:33Z) began after the obligation landed at 14:38:40Z; two more (25ee3009 13:05→18:43Z, d116c1ff 11:46→16:08Z) span the moment; the remaining SEVEN ended entirely before it. Stronger still: `grep -rl` over `~/.claude` finds the obligation's text in NO installed tree — not `~/.claude/rules/token-efficiency.md` (mtime 2026-08-25), not the marketplace plugin copy — only in two transcripts, one of which is the session that authored it. The only delivered copy is the main checkout's own projection (mtime 2026-08-30 16:39), reachable on demand via `load_context` and not otherwise. So the count of sessions that could have RECEIVED the obligation is not 'fewer than ten'; on the evidence available it is at most the one that started after 14:38:40Z, and possibly zero.",
+                "shape_addition": "`tool_using_requests` is recorded here and is NOT in the first entry's corpus block. It is added because without it the ratio is not reproducible from the entry — and the first entry demonstrates the cost: its `tool_calls: 3031` over the 2,889 tool-using requests its own § E2 reports gives 1.05, not the 1.01 it registers. The two figures came from different runs. A corpus block exists so a reader can recompute the ratio; this one can be recomputed (3266 / 3237 = 1.008959 → 1.01)."
+            },
+            "reading": {
+                "calls_per_request": 93.74,
+                "mean_batch_size": 1.01,
+                "blocking_share": 0.6641,
+                "context_floor_max": 226528
+            },
+            "delta_vs_baseline": {
+                "mean_batch_size": "NULL — 1.01 → 1.01, the pre-committed outcome. Unrounded 1.008959 against a first reading whose unrounded value was not recorded. The finer signal moved the same way and no further: multi-block requests were 27 of 2,889 (0.93 %) and are now 26 of 3,237 (0.80 %) — 24 of size 2, one of size 3, one of size 4. The step pre-committed to this being the RESULT, not a reason to repeat the reminder more loudly; that pre-commitment binds even harder here, because the corpus could not have received the reminder in the first place.",
+                "blocking_share": "+0.0439 — 0.6202 → 0.6641, which the gate reports as a REGRESSION and is why the run exits 1. DELIBERATELY NOT RE-BASELINED. It is not this step's metric, raising a baseline needs its own reason in its own change, and this is one local mtime-window reading.",
+                "context_floor_max": "-4,177 — 230,705 → 226,528, within the baseline in the gated direction (DOWN). DELIBERATELY NOT LOWERED: a ratchet is never tightened on a single local reading.",
+                "calls_per_request": "72.67 → 93.74. REPORTED ONLY, and the movement is the denominator again exactly as `directions` records: `api_calls` is comparable (3,052 → 3,281) while `user_requests` fell 42 → 35. Evidence of nothing."
+            },
+            "gate_verdict": "exit 1, on `blocking_share` alone. `mean_batch_size` passed."
+        }
+    ]
+}