Slice 3 of the completion wave: closing a contract that is a stage of a persisted Programme marks that stage COMPLETE, persists the ledger, and rides the continuation canary on the close envelope — the next eligible stage is SURFACED, never guessed (task done ≠ programme done). No programme bound → no continuation block; ledger failure degrades without failing the closure. Three contracts. Also fixed en route: contract ids were second-granularity timestamps that collided when two needs filed in the same second (red test proves distinct ids now, uuid suffix added).
The command seam validated intents but nothing EXECUTED them. dispatch() binds the canonical store, runs the validator, then executes accepted intents through the Construction Interface (resume/transition/amend_scope), refusing via the lifecycle with accepted=False+success=True (a refusal is a valid protocol outcome, not an error). Idempotent replay returns the recorded intent without re-executing. Three contracts: end-to-end phase movement, refusal shape, replay-no-double-execute; seam file 12/12. Slice 1 of 5 was stage_guards (#2467 wire wave).
The five orphaned V1 components: slice 1 wires stage_guards for real. transition(verifying|closed) now runs §15F readiness StageChecks derived from guard: verification obligations (or inline guard_checks); a blocked guard = VERIFICATION_FAILED envelope and the phase does NOT move — builder cannot self-certify past a failing check. building/free edges unchanged; capabilities() schema documents the new arg; CLI transport carries guard_checks as JSON arg 4. Four contracts red-first (blocked/pass/inline-override/free-edge); 56 interface+guards tests green.
Five separate dim system messages after every task/build response (budget, claims, growth, self-improve, plus the cron/queue echoes) spammed chat. New collect_post_response_signals() is a PURE headless collector with degrade-not-crash try blocks; the worker renders ONE line or silence. 6 frozen contracts: compose order [chat,task], resize semantics re-keyed, collector all-signals/silence/never-raises/cycle-multiple. 38 TUI/layout tests green; pyright error count unchanged at baseline 8.
Nick: chat-first feels natural. Compose yield order reversed (rune_tui.py compose), resize_left/resize_right actions re-keyed to the new visual positions (left=chat shrink now grows task), focus actions and all name-keyed query_one calls untouched (position-independent by design). Static source proof: yield order [chat, task]; 45 layout/TUI tests green.
Nick caught the deviation: module verdicts were substring-guessed from names (memory→rust, gitea→go) with a silent rust default. Authority order now: the measured classification map (import-evidence) decides FIRST for every module; hint hits are labeled "(name hint)" so provenance is visible; unmeasured+unhinted modules are "unclassified" (12), never a guessed rust. Store rebuilt: 106 rust / 45 go measured, 33+7 hint-labeled, 23 test-infra, 12 honest unclassified. Battery 95 green (test updated to pin the authority contract).
CI red at d67fc46: spine_language.py crossed the new-module henxel with zero coverage. Eight contracts: measured-map read + graceful degrade, kind-default verdicts, TUI/hints, subsystem member-majority, unqualified-discovery honesty (candidate stays DECLARED), legacy-ID normalization with edge re-pointing, and the closed-contract-only gather law. 8/8 green. Also: construction_commands dispatch WIP stashed on its own branch (slice 2, next).
The W0 map had drifted (174 vs 180 files): cc_loader/sc_loader/mr_loader/mems_loader (FFI kernel membranes → RUST), spine_language (deterministic carve → RUST), __init__ (metadata reads) now carry evidence-backed side records and measured function/line counts. spine_language._classification_sides() — the runtime reader — sees 180 entries; spine battery green.
The 0.99.0 bump ran CI against the delegation code before the annotations landed: BLE001 345->349, S110 75->81 (my new kernel-failure fallbacks), plus spine_language unmapped in the umbrella manifest (test_umbrella_impact) and the resulting spine-rebuild membership diff (rebuild-neutrality). All four new fallbacks carry intentional-silence noqa annotations per the ratchet rule (count now exactly 345/75 — baseline UNCHANGED, ledger honest); spine_language mapped into perception next to spine; store rebuilt so the membership edge and language stamps are in the committed store. 127 tests green incl. the four CI-red surfaces.
The W3.4 kernels existed but the Python publics still ran Python. Now: witness extract_topic/detect_signals delegate through the _stopwords()/_tables() gather seams (PATTERN-TABLE GATHER-SEAM LAW) with _py_ oracles; abstraction extract_facts gains the deterministic kernel FALLBACK when the LLM returns nothing; tier_manager was already wired; the RUNE_COGNITION_CORE=python pin is now honored by a SHARED cc_loader.core() probe (the four module-local probes silently ignored it). The canary matrix caught a REAL kernel bug: regex_lite dropped \b anchors as literal b chars so every word-boundary row silently missed (bwrongb never appears in text); rewritten as alternatives-expansion + segment-variant DFS gap-walk; 34-row matrix judged against REAL Python re shows 0 mismatches. canary_growth_pins 27/27 x5 both engine modes; canary_unbounded 14/14; per-row fixtures carry re-truth, not hand labels (a lying fixture is worse than uncovered).
spine.py crossed the 1200-line giant-file debt cap (1203) with the language-verdict addition — CI lint red. The three transforms (language placement, discovery-ID normalization, contract-discovery write-back + gatherer) move to spine_language.py intact, re-exported through spine for import-compat (structure unchanged: 256 entities, same verdicts; spine battery 87, canaries 19/19 + 21/21, henxels hold, spine.py back to 1034).
The kernels and FFI landed in b0a682b but the Python publics still ran Python — the strangler breach this slice owed. Live probe caught the FFI-arg-shape defect: promote/demote take the FULL ENTRY (kernel recomputes current tier internally), not a tier string; passing a string silently returned threshold-default verdicts (promote cold went warm-to-hot WRONGLY via defaults, demote rock went cold ignoring rock protection). 12-case kernel==oracle sweep green: wash>=20, recency hot/warm/cold, naive-timestamp parity trap, rock protection. PY oracle bodies kept; RUNE_COGNITION_CORE=python pins fallback; canary_growth_pins 26/26 both engines; canary_unbounded 14/14.
The five unassessed subsystems are containers: their language derives from member modules measured sides (classification.json, the W0 dichotomy map, rust and go counts carried in the verdict text). Unclassified members noted rather than guessed; unreadable classification degrades to unassessed. Spine canaries 19/19 and 21/21, spine battery 87 green.
Replace the obsolete removeLabel(name) pattern assertion with the actual Gitea /labels/{id} request, check ID lookup under the 404 idempotency guard, and require real P0–P3 managed labels. Local canary 39/39; workflow JS regression passes.
Issue #2467 carried both P0 and P2 because name-based label DELETE returned 404 and the workflow silently ignored it. Use the numeric ID API, enforce exactly one priority, clear stale unprioritized flags, and execute the workflow JavaScript against a Gitea-compatible fake in regression tests.
Migrate 22 malformed contract-discovery IDs to construction-surface, retain evidence/edges, keep candidates DECLARED, unify cold/warm language projection, and refresh the pinned aiohttp lock. Red regression tests precede the fix; local full suite 2301 passed, spine canaries 19/19 and 21/21.
Contract BUILDER-2026-10-05T12:03:34Z-cli — FIRST SLICE under the 90/100
verification law (64172d8).
Kernels (cognition_core/broker.rs):
- tier_get/promote/demote: thresholds+now injected (no-clock); wash>=20
gate; rock immobility on demote; naive/malformed → designed fallback
- witness: extract_topic (word-run + stopwords-as-data), detect_signals
(error/correction/completion/topic-drift/commitment/contradiction/
preference/urgency/learning; pattern tables travel as data; the regex-
lite subset is faithful to the tables' shapes)
- safety: get_params projection, update_pattern bump/new-row semantics
- abstraction: extract_facts sentence/content-word projection
- crate-level probe tests (parse_probe.rs: aware parse, epoch, tier recency)
Protocol upgrade (issue #2465) — three additions to the machine itself:
1. PER-CONTRACT MINI-DOC: construction_projection.mini_doc() emits the
LLM quick-lookup page (what was built / connections / how to update)
per closed contract; close() writes it to store/scaffold/docs/<cid>.md
2. SPINE WRITE-BACK: spine.build() now consumes closed contracts'
discovered entities (expected_new_observed + candidate corrections) —
the spine learns what construction actually built (256 entities now)
3. LANGUAGE VERDICTS: every spine entity carries identities.language
(198 rust / 31 go / 6 tui / 5 unassessed); other projects get the
fit assessment at classification
Also: one lost FFI block restored (the stash dance dropped
cc_tier_*/cc_witness_*/cc_safety_*/cc_abstraction_* — the battery
caught it: undefined symbol).
Gates: canary_growth_pins 26/26 x3, persona 18/18, cognition_core
38/38, suite 2299/0, ruff clean, type ratchet PASS, henxels all-hold.
Nick's catch (2026-10-05): the auto-label bot highlighted that issues
carry priority only in their titles — the PRIORITY GATE job demanded
P0-P3 labels but nothing ever ASSIGNED them programmatically. The title
inference even used a dead label ('priority: high', normalized away
2026-09-04).
- auto-label's priority inference re-implemented in the gate's own
namespace: title prefix \bP([0-3])\b first, then keyword inference
(spine/critical/urgent -> P0; broken/fix/repair/defect -> P1;
kernel/port/refactor/improve -> P2; benchmark/measurement/deferred ->
P3); managedLabels now carry the REAL P-labels
- paired with the custodian fix: create_issue assigns the P-label at
FILING (no issue ever lands unprioritized)
- full open-issue set backfilled (2461-2466, 2455/2459) and the live
probe (#2466, self-verifying) closed after verifying P1 landed
at filing
- 100% branch coverage on the deterministic Rust kernels (achievable,
exhaustive via parity matrices)
- ≥90% per-slice changed-lines working bar, aiming higher per slice
- floor ratchet climbs after clean waves (70->76->84->90)
- honest fixtures only — a lying-fixture branch is worse than uncovered
- assertion quality counts: value-verifying asserts only
Missing from the previous commit's staged set (the harness caught it):
the full W3.3 kernel block now lands. All parity walls verified:
canary_growth_pins 9/9, canary_persona 18/18, canary_cognition_core
38/38, suite 2298/0, diff-cover 72%.
- identity_anchors: check_drift delegates (pattern table stays the
module's data authority via _pattern_triples); _py_check_drift oracle
covers the four prefix families; drift parity proven on the full
table (31 patterns) + the full-text multi-class case
- growth_delta: _near_dup/growth_profile delegate; oracle accepts delta
dicts AND raw strings; near-dup kernel corrected to the REAL
semantics: prof-key = strip-punct lowercase words >4 chars, split at
|A∩B|/max(|A|,|B|) >= 0.25 (NOT a jaccard-min heuristic)
- pins: recall kernel projects via _project parity (details-blob parse,
defaults, 'reason' field); empty-context abstention frozen
- probe branches exercised honestly: pinned-fallback + ImportError via
the from-import resolution path (the from-import goes through the
package attribute, not sys.modules — the PEP resolution law frozen in
the battery)
- reflect malformed-JSON + [:3] cap + apply budget cut frozen
- suite 2295/0; diff-cover 72% (floor 70); canaries 9/9+18/18+38/38
x3; henxels hold; ruff/type clean
Nick
conductor