l-gevity/l-gevity-skills v3.6.0

on GitHub

Lighter minimal path, placement records that conform

This release makes the routes an agent takes most often cheaper to load, without changing what they decide. Every change was measured with recorded scenario runs; where a cut was in doubt, it was tested three runs against three. One cut was reverted when its first A/B looked worse. It was retried after fixing the contradictions that A/B exposed, and kept when the outcomes matched.

Routing

• alchemy is now a dispatch core with aliases, dispatch, gate prerequisites, companions, and the output contract (4,375 to about 1,900 words). The qualification route, hand-offs, handshakes, checklist, and retrospective order moved to references/adaptive-route.md, which only ADAPTIVE, FULL, and audit read. Scenario checks prove skips and single-gate runs never read it and structural runs do.
• Each stage loads when the route reaches it. A blocking hand-off decision ends that candidate's route.
• Every stage that ran emits its own record after the trail. Multi-gate runs had summarised the placement gate only in the trail table.
• PROVISIONAL grounding passes with its confirmation queue. Readiness then decides whether the unconfirmed requirements can change the increment. The hand-off table and design diagram had called it blocking; the stop rule and every recorded run had not.
• The stack skill (-javascript / -python) loads only when rule files are written or changed.
• CLAUDE.md section 6 keeps only what every session holds. Gate 3, handshake, two-pass, and ownership mechanics live in the router and the gate skills.

Skills

• morphogenetic-architecture: split into a Rapid core (4,778 to 2,764 words) and references/full-analysis.md, read on entering Full. On a 3-vs-3 A/B the outcomes matched: every expectation met, every placement record conforming, the same decisions. A quick placement loads 29 KB of skill text instead of 43 KB. Rapid runs must read rapid-topology-scan.md and position-legality.md; a third had read SKILL.md alone.
• Two contradictions fixed in the placement skill. The rapid scan told Rapid to emit restructuring fields as Not required, and §3 asked a cross-domain PLACE for a Prediction; §8 and the report checker said both omit them. The template now lists its 14 core lines before the 7 restructuring lines, and a PLACE, KEEP, or DECLARE-RUNTIME-CYCLE record ends at Verification. Records passing check_topology_report.py went from about half to nearly all.
• functionality-complexity-tradeoff: restatement removed, section numbers kept (4,411 to 3,878 words).
• Path-conditional sections moved verbatim to references: the assembler in architecture-as-code and architecture-as-code-javascript, grounding's Recovery Mode, and simplification's reduction-operations catalogue.

Library enforcement

• New scenario checks: reads_include, reads_exclude, and topology_records, which runs the repository's report checker on every placement record. The not-run check reads a record field's wrapped continuation lines, so "X considered, / not selected" no longer counts as a claim.
• Twelve scenarios and 65 expectations. 63 are met; 2 are listed as known failures with their observed rates: one PLACE record carrying restructuring lines (1 of 13 since the template fix), and one probationary MOVE that omits why its evidence cannot be measured. New scenarios: placement-rapid-bounded, placement-escalates-to-full, and route-stops-at-blocking-worth. morphogenetic-architecture is now covered.
• The mutation test catches 49 deliberate violations.

Known limits

• A structural run still loads about 210 KB. The route requires ten stages for a new subsystem, and most of each skill is the method itself. Making it much lighter means running fewer stages, which would change what a structural run guarantees.
• Most committed transcripts are single runs; the before/after comparisons rest on three runs per side.

Upgrading

Pin this revision with L_GEVITY_SKILLS_REF=v3.6.0.

Release commit 809ab39, covering 0f2b337, c870c76, and 3052571.