Developer preview. Pre-admission. Not for sensitive, proprietary, regulated, or production code. Expect breaking changes and data loss. Read the limits.
Bridlewise
GitHub

Unwired findings — the single ranked list

Date: 2026-07-27 · Status: research only, nothing edited or wired.

This is the one list. Two sweeps and three verification passes produced two evidence documents; this file is the authority over both, and exists because leaving two documents answering one question is the same “one question, two paths” defect the doctrine forbids.

Two conclusions the evidence documents establish before any item below matters:

  1. The 57 sibling workspaces hold almost nothing. Total salvage from all of them: one 25-line mutex, one uncommitted test-classification idea, one research document. The real mass is inside harness-main-v1.
  2. No experiment anywhere contains work newer than live’s version of a file they share. Measured over 79,694 same-path content comparisons: zero. So the only genuine gaps are things absent from live, never things stranded in a snapshot.

Tier 1 — do these

1. The exposure catalog is fabricated

10 of 11 modules in integration/src/shared/exposure/registrations/ are per-process mock state. Their own comments say so (tools.ts: “mirrors what the real tool registry would expose”). Zero writers anywhere. privacy.ts is the lone real one and is the template. Live via routes/schema.ts → /api/schema/ui/values → the client. This is fabricated data on a user-facing path — the first prohibition. Either derive them or delete them; shipping invented numbers is the one option that is not defensible.

2. EXP429 — 306 additive lines into symbol-search.ts

diff live exp429 = 0 removed, 306 added. A literal drop-in adding BM25, query specificity, hybridSearchSymbols, bias detection. Live’s memory-search.ts header advertises the plug-in point this was written for. ~1 hour, and it may make item 3 unnecessary. experiments/exp429-improve-core/src/intelligence/symbol-search.ts

Verified directly, not taken from the agent’s report (this is the top recommendation, so its central claim should not rest on someone else’s diff): diff of live against the experiment gives 0 lines present only in live and 306 present only in the experiment, and every symbol live exports is still exported by the experiment version. The additive-superset claim is exact. What remains unverified is the quality number (88-100% top-1) — that is self-reported and nothing was executed.

3. search_code is a recursive regex grep

web-interface-foundation-v1/server/tool-executor.ts:335 — readdirSync walk plus regex.test(line). No index, no ranking, no natural language. EXP398’s recorded decision was to replace it with BM25 + re-ranker; the other half of that same decision landed (quick_view, slice, viewport, skeleton are live). The model’s only way to find code by description is a regex it has to guess. Do item 2 first.

4. Landlock never reaches the web UI

experiments/exp486-landlock-folder/server/landlock-runner.ts — 475 lines, 15 tests, designed, with the guardian it depends on already ADMIT-reviewed twice and byte-identical to live. grep -rniE "landlock|guardian" across the web server and cockpit returns zero. Highest-confidence, lowest-risk item in the inventory. Small-medium.

Tier 2 — real capability, real cost

5. Trace mode (USR-057) is two-thirds built, and both governing documents deny it

integration/src/performance/ — 8 modules, ~65 KB, landed 2026-07-24. metrics-collector.ts and performance-config.ts are built and wired (server/metrics-bridge.ts:33), with the config carrying the exact USR-057 key list. The gaps are narrow and specific:

6. The harness cannot tell whether the model did what it said

Plan-adherence: parse the plan the model states in prose, compare to the tool calls it actually emits, classify divergence four ways, feed a correction back. Two SHIP-IT experiments, 100% drift detection, 18/18 tests, integration already designed against a session loop. grep -rli adherence integration/src → zero. Live’s deviation-detector answers a different question (“did it wander off-topic?”). Caveat: its tool vocabulary is CLI-era and must be re-derived from the live registry. experiments/llm-adherence-live/, experiments/adherence-automatic/

7. tool-call-gate.ts — a capability regression inside a wired module

Live is a 297-line rewrite of a 792-line version; live’s header still reads (EXP485, ) with a dangling comma. Missing: invalid_json, all_blocked incoherence, pluggable rules, strict extra-args, structured schema issues, an exposure surface, and Levenshtein “did you mean?” — which turns a dead-end block into a self-correcting one. Port the behaviours onto the live file; do not blind-replace.

8. _experiments/llm-beyond-max/ — 2,985 lines, no live counterpart

Chunking, continuation, map-reduce, sliding window. Live handling of the same concern is one line surfacing finish_reason === "length" as an error. Only matters if you want outputs longer than one model response — but there is no alternative in the tree.

9. Multi-agent orchestration and MCP have no web surface

integration/src/agents/** (2,793 lines, 464 tests) — bridge is 50 lines that only lists agents; the four delegation tools are absent from the web manifest. integration/src/mcp/** (20 modules, 161 tests) — bridge is 78 lines: “Does NOT connect to servers.” The UI lists servers it cannot reach.

Tier 3 — cheap, unglamorous, worth doing

Delete, do not wire

The finding that is not a module

The mandated entry-point document is stale in the direction that suppresses search. integration/AGENTS.md says “read this file completely before taking any action”, and:

Claim Measured
~323 modules registered 526 manifest rows / 509 files
~6,949 tests across ~229 files 403 test files
10 open-blocking findings 12
performance observability: “research not started” 8 modules, one wired

Every agent starts from a picture wrong in the same direction — it under-reports what exists. An agent told there is no performance work will not go looking for function-manifest.ts. That is very likely why USR-057 was carried as unstarted while two-thirds built. All four numbers are derivable, so a generated status block compared on every check run turns this from something someone must notice into something that fails a gate — the same move function-manifest.ts already implements for functions, and is never called.

What no pass would have caught

Three passes ran: path-set difference, file size, and tree-wide exported-symbol names. Two false finds were produced and killed, both the same shape — a symbol absent from the path it used to occupy but present elsewhere in live (createMcpToolAdapter → mcp/tool-adapter.ts; proactiveRetrieve → proactivelyInject). Relocation is the normal outcome of a refactor here, so a name-grep scoped to an old path manufactures findings.

The fourth pass closed that blind spot

The three passes above compare by path, by size, and by symbol name. I claimed a capability renamed and moved and rewritten in one change would escape all of them. A fourth pass was run to test that claim, comparing content rather than identity: every experiment src path with no live counterpart was scored by normalized line-set Jaccard against all 509 live files, and the best match reported.

It works, and it caught exactly the case I said it would miss — memory/unified/inverted-index.ts matches live memory/unified/unified-inverted-index.ts at 0.96: renamed, and still found. Others resolve at 1.00, i.e. an identical body under a different path:

Experiment path (flagged “absent”) Live file Overlap
unified/awareness/signals/risk.ts awareness/signals/risk.ts 1.00
unified/awareness/signals/trending.ts awareness/signals/trending.ts 1.00
quality-packs/multi-lang/checks/complex-condition.ts quality/checks/complex-condition.ts 1.00
quality-packs/multi-lang/checks/duplicate-imports.ts quality/checks/duplicate-imports.ts 1.00
memory/unified/inverted-index.ts memory/unified/unified-inverted-index.ts 0.96

This corrects a number in the evidence. Sweep 1 reported “335 distinct src/**.ts paths absent from live”. That figure counts relocations as absences. Measured by content, the genuinely-absent set is 259 — and the relocated remainder independently confirms two “delete, do not wire” verdicts above, since _experiments/unified/ and quality-packs/multi-lang/ turn out to be the same bodies live already carries, not merely the same concerns.

What is still not closed

Nothing was executed. Every performance figure quoted from an experiment (EXP429’s 88–100% top-1, EXP460’s 71.4%, EXP449’s 97 tps) is self-reported by that experiment and has not been reproduced. Believing the ranking above means believing those numbers; confirming any single item means running it. That is the honest boundary of a read-only sweep, and no further static pass can move it.

2026-07-29 — DECLARED-BUT-UNREACHED, found by the silent-universe work

The unifying shape: a declaration with nothing behind it, where the guard cannot tell.

# declared reached how it stayed green
U-01 RoomId included "memory" and "tools" no MemoryRoom.svelte / ToolsRoom.svelte all 9 consumer tests stubbed roomExists/outletExists; the real glob was never asserted non-empty. 2 of 7 rooms permanently dead. FIXED (S16) — removed at rung 2, openRoom(id: RoomId) made the empty state representable
U-02 tests/e2e/ in playwright.config.ts testMatch the directory does not exist all 5 specs live in tests/cockpit/e2e/ and matched only because e2e is a bare SUBSTRING of the surviving alternative. FIXED (S17) — anchored at ^, and its first anchor (?:^|/)e2e/ still matched inside cockpit/e2e and was caught by a plant
U-03 ToolResultStore.configure zero callers it assigned into a deep-frozen object, so it would have thrown for every caller on every path. FIXED (S14) — removed
U-04 JudgeOptions.budgetMs declared, passed by every caller, read by nothing FIXED (S13) — removed; 0 occurrences across 975 web source files afterwards
U-05 integration/src/self-improving/ — 9 files reached by NOTHING, runtime OR type level, from 281 entry points no consumer-count rule sees it, and it collides by name with web/server/self-improving.ts. OPEN — dead-domain finding for the user
U-06 stripCommentsKeepStrings byte-identical to stripCommentsOnly two names for one function; 11 callers. FIXED (S17) — callers moved, function gone
U-07 OUTLET_IDS (room-door-owner.test.ts:77) a hand-written literal typed readonly OutletId[] a subset array is assignable, so a seventh outlet with a dead glob gives 6 passed, 0 tsc errors. OPEN — the guard built to end U-01 contains U-01
U-08 webUiTestRuntimeBudgetMs replaced FIXED (S13) — 0 hits repo-wide after

The rule these produce → contract §31b: derive the universe from the declaration; a guard that lists its own universe by hand contains the defect it exists to end.

2026-07-29 (later) — GUARDS THAT WERE UNWIRED FROM THEIR OWN SUBJECT

# guard what it could not see
U-09 frontend-contracts-owner §D matched an import specifier against a stripper that blanks string interiors — 5 tests passing unconditionally
U-10 static-literal-extraction-cache assert.equal(new Set([...aSet]).size, arr.length) — true by construction; a planted duplicate is certified unique
U-11 dead-export-scanner 0 dead of 2,833 — the file is inside its own reference surface and its 315 baseline strings held every baselined export alive. Removing the baseline: 0 → 124; excluding itself: 127
U-12 same 140 of 315 baseline entries MALFORMED (whole source lines from a listing and a diff), landing in a set discarded via void missingBaseline
U-13 filesystem-owner:641 etc. no floor at all — rename one symbol and the loop enumerates nothing, body never runs, green in 0 ms
U-14 declared-universe reason hatch no member-count check of any kind — 7 planted exemptions all silenced it
U-15 room-door-owner OUTLET_IDS a hand-written universe typed readonly OutletId[] — a subset array stays assignable, so a 7th member arrives with no floor
U-16 the .gitignore guard its own REMEDY TEXT printed a pattern that reproduces the defect it rejects
U-17 shared-primitive-bindable-contract:162 globs src/lib/shared/*.svelte — after the cutover it matches an empty directory: zero tests generated, green. OPEN — must be floored BEFORE the cutover runs

The shape: a guard whose universe comes from the tree can be unwired by the tree. §31b (derive the universe), §31 (a partial loss is invisible), and the new door rule — a count is the wrong floor, because zero is often correct; what a count cannot separate is whether the TERM still means anything.

2026-07-29 (S24) — U-17 REFUTED, and the finding underneath it

U-17 as recorded is WRONG, and it was wrong about the mechanism, not just the severity. It said shared-primitive-bindable-contract:162 globs src/lib/shared/*.svelte so that “after the cutover it matches an empty directory: zero tests generated, green.” Driven in both post-cutover states:

state result
src/lib/shared/ absent rc 1, ENOENT scandir at collection, Tests no tests
dir present, .svelte gone from disk AND index rc 1, 3 failed — all three floors fired

The cases were generated by readdirSync (line 72), not by the glob; the glob only resolved module paths for loading. The file already carried three non-emptiness floors. It could not pass vacuously. U-17 is CLOSED as REFUTED, not as fixed.

# guard what it could not see
U-18 shared-primitive-bindable-contract a LIVE member of its own class, permanently. $bindable is declared in 2 of 150 tracked .svelte files; the guard enumerated a DIRECTORY and could only ever see one. src/cockpit/shell/Desk.svelte (text = $bindable(), consumed by Cockpit.svelte) is in the UI that survives the cutover and was never covered. FIXED (S24) — retargeted at the component tree
U-19 same, its RULE the rule was fitted to the single instance it could see: it matched bind:{prop}, conflating the DIRECTIVE NAME with the BOUND EXPRESSION. It passed only because SearchInput’s prop is value and its element’s directive is also value. Desk.svelte’s correct bind:value={text} is REJECTED by it. FIXED (S24)
U-20 the ten in the sweep below generate cases from an enumeration of the doomed tree with no non-emptiness floor at all. OPEN

THE SWEEP, and the sample was the population. 508 tracked test-tree files → 149 call an enumerator → 130 generate cases from it → 35 name the doomed src/lib, of which 10 have no floor. Spellings [UNIT=files]: walk 84, readdirSync 45, walkTs 20, sourceFiles 18, sourceFilesReachableFrom 5, ls-files 4, nonEmptySourceFilesUnder 3, gitTrackedPaths 3, sourceFilesMentioning 2, reachableFilesImporting 2, globSync 2, import.meta.glob 1, glob 1.

import.meta.glob — the spelling U-17 named — occurs in exactly ONE file in the whole test tree. The real population is walk and readdirSync. And walk’s assertIndexCoverage does not help here: after a real git rm the index agrees with disk and the empty walk is legitimate.

THE RESIDUAL NO RUNG CATCHES: a deleted specifier is TS2307; a deleted scan root is refused by walk(). import.meta.glob with a non-matching pattern is silent at every rung — not a specifier, not a scan root, and it yields {} rather than an error. That is why the floor must be on the SUBJECT and not on the count.

2026-07-29 (product round) — DECLARED-BUT-UNREACHED, and one DECLARED-BUT-UNSTYLED

# declared reached how it stayed green
U-18 .absent in RoomHost/ScopeHost/Desk the rule carried ONLY margin “nothing to inspect yet” was typeset identically to ordinary content. FIXED (S26)
U-19 the desk’s “reading the recorded trip” byte-identically styled to “carries nothing” pending and absence were indistinguishable. FIXED (S26)
U-20 the three-colour status scheme reached exactly ONE element in ONE room [data-state="ok"] .beat-card + 3 siblings; any other surface had to re-derive the hue. FIXED (S26)
U-21 GRAVITY-WELL’s motion vocabulary 9 of 9 tokens exist in NEITHER tokens.css re-measured independently by S28 (all nine, both files) plus the four conflicting axes. §7, §7b and the §11 src/styles/tokens.css line are now marked SUPERSEDED in the doc itself, with the axis table and the withdrawal. CLOSED (S28)
U-22 which tokens.css supersedes which no design document decides DECIDED in DESIGN-COCKPIT-V2.md §12 + §13 on a measurement nobody had taken — the CONSUMER CENSUS: src/styles/tokens.css has zero readers in src/cockpit/** and 225 distinct names read by src/lib/** (+7 by App.svelte), so its entire readership is the tree §13 deletes. Intersection re-measured at 0 (208 vs 230). CLOSED (S28)
U-23 --beam in the sweep canvas captured ONCE at mount fixed at src/cockpit/paint/design-tokens.ts — reader revoked when the paint ends, plus a NAMED signal (designConditions) the repaint binds to. The class was bigger than --beam: all 15 SweepPalette values are re-declared by the light theme, and three more canvases read fresh but re-ran on nothing. Sweep/MicroSweep/Spark/SwitchRoom converted; guard toEqual([]). CLOSED (S28)
U-24 enum setting options raw server values on screen — and it was FOUR sites, not one: the segmented control, the index row’s reading, the provenance line and the reset button derived through humanizeId, the id→words owner the project already had; SettingRow no longer carries a readonly string[]. The "short-term" specimen is not a served enum member — the real ones are round-robin, oldest, tfidf, fifo. CLOSED for the ~27 members whose token carries its words; the 3 acronyms need the WIRE to carry declared words → new row U-27
U-25 the motion half of #94 the prescribed mechanism does not exist built at src/cockpit/motion/motion.ts: a motion cannot start without a MotionCause, no factory takes a clock, becauseChanged REFUSES an unchanged pair, and a cause is spent when used. All 3 of 3 self-starting animation: declarations converted; guard is toEqual([]), not a ratchet. CLOSED (S28)
U-26 redactedValues success counter increments on a 3-of-48-character mask the redactor masks the NOUN, leaves the VALUE, and reports success. FIXED — S29 (47edef0: a redaction that did not remove the value can no longer report success) + S32 (the counter counts VALUES → U-30)
U-27 an enum member whose TOKEN does not carry its words (fifo, lifo, tfidf) no derivation over the token can invent them — humanizeId("fifo") is "Fifo", and nothing can tell fifo from oldest by inspection the words must come from the DECLARATION (makeEnumCustomization + the literal registrations beside it), which means the wire carrying words per member — a contract change across both repos. OPEN — S28-01
U-28 src/lib/intelligence/DependencyGraph.svelte + src/lib/timeline/HorizontalTimeline.svelte read design tokens and re-run on nothing — the same class as U-23 deliberately NOT fixed: both are in src/lib/**, the tree COCKPIT-V2 §13 deletes, and the tokens they read have no theme variant today. OPEN by decision — S28-02

2026-07-29 (security + gate round)

# declared reached how it stayed green
U-29 “a sensitive KEY NAME discards the whole value whatever it looks like” gated on typeof v === "string" arrays and nested objects kept their secrets, and both counters read exactly like the inert control. FIXED (S32)
U-30 redactedValues — “EVENTS in which at least one secret was… removed — a SUCCESS CLAIM” counted per SPAN one droppable key + one surviving secret incremented the success claim. FIXED — it now counts VALUES
U-31 both counters: “re-observed absent by the canonical oracle” the file imports no privacy module at all residue of a reverted wiring, left in a security control’s contract. FIXED — claim DELETED, and the honesty test uses the oracle as its own witness
U-32 enabledSinks / redaction.kinds declared in the contract, SET in the shipped config, READ BY NOTHING and DEBUG_SINK_KINDS declared file and timeline, neither implemented anywhere. enabledSinks now READ; the two unimplemented kinds REMOVED. redaction.kinds still unread — stated, not faked
U-33 VITEST_MODULE_CACHE_DIR a hardcoded constant naming the default directory a redirected run built its cache elsewhere while the reporter counted an untouched one — a fully cold run recorded as WARM. FIXED (S31)
U-34 set-cookie under the forbidden-pattern list matches none of the eight patterns never reaches the key-name decision. FIXED (S33) at the key-name gate — dropped in all four casings (T-check A: each +1/+0). The regex LIST itself survives as performance.labelSchema.forbiddenLabelPatterns → U-36
U-35 the absence-witness classifier accepts a constant, an unused const, a different subject, and a witness that passes with the subject EMPTY §36’s own failure mode inside the guard enforcing §36. Ratchet 28 → 21 after a missing presence spelling was added — it had been charging debt for compliance. LINKAGE still OPEN and recorded in source

2026-07-30 (S33 triple-check — two independent read-only checkers)

# declared reached how it stayed green
U-36 “a single-casing pattern is UNREPRESENTABLE” (S33) true only for debug.redaction.additionalSensitiveKeyNames — one subsystem over, performance.labelSchema.forbiddenLabelPatterns still holds 13 single-casing regexes (bridlewise.config.json:580) behind plain z.string().min(1).max(512) (performance-config.ts:90) and the one list is compiled TWO WAYS: isLabelAllowed (performance-config.ts:270) compiles "u" — [Ss]ecret cannot match SECRET, so driven, SECRET/TOKEN/API_KEY/PASSWORD/AUTHORIZATION/SK- values were ALL ALLOWED through the gate — while redactLabelValue (:282) compiles the same list "giu" and drops all seven. The two gates disagree about SCREAMING credential material; any consumer of isLabelAllowed alone admits it. Found INDEPENDENTLY by both checkers. FIXED (S34 2c3a844): field → forbiddenLabelNames (names only, .regex refuses — all 13 shipped patterns fail config load, driven); ONE predicate isForbiddenLabelText serves both gates (a second compilation site has nothing left to compile), backed by a one-interpreting-reader tree guard; the credential half is ALWAYS-ON from the declared owner, so emptying the config field cannot re-admit credentials. 9 mutations, 9 killed
U-37 DATABASE_URL / SK_ survive in every casing (all four rows +0/+0) — shape-invariance proves a VOCABULARY gap in SENSITIVE_KEY_PARTS (no part matches databaseurl/sk), not casing closing it from config was DECLINED (§37): the config field covers one of four consumers, restoring exactly the two-owners drift secret-signals.ts exists to end. HALF-CLOSED (S34, precision corrected by its check): SK_ now caught in the four isSensitiveKeyName consumers — debug gate, perf label gates, web redactor, web write gate, ALL FOUR DRIVEN — via the owner pattern /(?:^|[^\p{L}\p{N}])sk[_-]/iu; the egress classifier (classifyData) is parts-only and does NOT catch it (→ U-42, driven: SK_TEST → public, signals=[]), nor does web’s harnessOwnSecrets (→ U-44) — separator-decided, so it cannot match disk_/task- (both pinned as negative fixtures; the old [Ss]k[-_] bigram had been redacting them). Residual is DATABASE_URL ALONE, mechanism recorded: databaseUrl is a legitimate writable config-field name, so classifying the NAME sensitive would make the web write gate refuse it — the trade belongs to that gate’s owner. OPEN
U-38 no-temp-dir-leaks.test.ts:75 — “aliasing cannot evade it” the import-edge branch is [^\n]*-bounded witnessed: single-line import { tmpdir as t } from "node:os" detected true, the multi-line aliased form detected false — and the t() call site never matches \btmpdir\b. The same \n-exclusion class S33 fixed in importSpecifiers. FIXED (S34 37573cd): bound [^\n]* → [^;]* (an import clause can span lines but never contains ;); the guard now reads through the sanctioned SourceView — a NEW view codeWithStringsAndTemplates(), because codeWithStrings() would have gone silently blind to hardcoded `/tmp/…` template interiors (a shrink the agent refused). Class-swept: 26 + 33 candidate lines, no further import-scan member
U-39 importedNeighbours (module-load-order.test.ts:91) derives neighbours only from from "…" clauses a side-effect import "./x.ts" or a dynamic import would not join the 21-entry population. None exist in the guarded subtrees today (positive witness: grep finds only comment text). OPEN — latent, filed

2026-07-30 (S34 round — found and FILED, not fixed; each with its mechanism reason)

# declared reached how it stayed green
U-40 token-prefix value detector (secret-signals.ts:74) compiles without i — an uppercase SK_AAAA… (20+ tail) VALUE carries no signal from that detector the key-name gates now cover the common case, but the data classifier’s VALUE channel doesn’t; admission semantics reach the web config write gate, so it is an owner-vocabulary decision. OPEN — filed (S34-01)
U-41 the secret-shape catalog vs connection strings postgres://user:password@host matches NOTHING in SECRET_TEXT_SIGNAL_CATALOG — no URL-userinfo detector exists a DATABASE_URL VALUE has no shape signal at all, compounding U-37’s name gap. OPEN — filed (S34-02)
U-42 SENSITIVE_KEY_PATTERNS (the auth word rule, now the sk- prefix) classifyData (data-policy.ts:44) consumes normalizedKeyHasSensitivePart ONLY — the patterns never apply to the egress classifier’s key gate that path deliberately matches DECODED keys while the patterns are documented as needing the key “AS WRITTEN” — reconciling the two views is a semantics decision, not a one-liner. FIXED (S36 722fe2d): the decode stays the classifier’s duty, the owner is asked the WHOLE question on decoded AND raw spellings — and normalizedKeyHasSensitivePart is now MODULE-PRIVATE, so the narrower question cannot be asked from outside the owner. Probe folded into S1-B05-CONTROL (the spec’s closed world is three scenario kinds — folded, not widened); mutation killed
U-43 neverPush.additiveRules[].pattern (git-adapter.ts:1012, config:491) compiled "u" without i in a security DENY gate: \.(env\|key\|pem\|p12\|pfx\|keystore)$ misses .PEM/.ENV one compilation site (not a U-36 member), but single-casing-adjacent; case-folding PATHS has its own false-positive surface, so the decision is filed rather than assumed. OPEN — filed (S34-04). Also the next candidate for the one-interpreting-reader guard (T1-S34-W1)
U-44 web harnessOwnSecrets (server/config.ts:1112) — env-credential discovery a FIFTH parts-only consumer — driven: of {SK_TEST, harmless_note, SECRET_THING} only SECRET_THING is discovered as an own-credential same class as U-42/U-37-residual; an env var named SK_… is invisible to the harness’s own-secret discovery. FIXED (S36 web b32c2f9): discovery asks isSensitiveKeyName on the AS-WRITTEN name and the local normalization re-spelling is deleted; first-ever direct test of the function, driven with a synthetic env, mutation killed — after the first fixture (SK_LIVE_TOKEN, whose token satisfied the parts half) SURVIVED the mutation and was replaced (SK_LIVE_9Q7)
U-45 the one-interpreting-reader guard (S34’s rung-5 backstop) matches \bforbiddenLabelNames\b in the code() view only — driven: both config.labelSchema["forbiddenLabelNames"] and "forbidden"+"LabelNames" pass the guard GREEN catches the natural identifier spelling, not a deliberately string-keyed reader. A limit of the rung, recorded; the T1-S34-W1 dynamic discoverer should close it for the class. OPEN — filed (S34-check)

The logbook index

Search the record

Type to search the curated record.

↑↓ move Enter open Esc close