qrspi-plus v0.7.3 — pipeline-correctness fixes + active-skill footprint reduction

2026-06-15 — v0.7.3: pipeline-correctness fixes + active-skill footprint reduction

Closes the eight P0 defects surfaced by the v0.7.2 self-host run plus a release-wide active-skill-prompt footprint reduction, shipped in a single phase with a single end-to-end slice. 45 tasks across 1 phase / 7 waves; 192 commits.

Goals shipped (G1–G9):

• G1 — verifier sidecar grounding: pre-existing anchor disambiguates work-unit-external vs prior-round code; rubric band 72–78 hardening-relevant correctness fixes now reach the keep set.
• G2 — bats @test description hygiene: [Tnn] and R\d+-F\d+ tokens swept from the corpus + permanent CI lint (tests/lint/test-bats-test-name-id-hygiene.bats).
• G3 — design absorption marker pipeline: explicit ## G<n> — <name>: absorbed by CD-<n> headings in design.md drive scripts/design-absorption-markers.sh → scripts/review-prep.sh plan-step TSV, eliminating implicit traceability gaps.
• G4 — plan-step upstream-paths: new scripts/upstream-paths.sh resolves the design absorption map to plan-task upstream pointers (closes silent-fan-out class).
• G5 — orchestration boundary observability: new scripts/orchestration-boundary-check.sh runs at phase end, writes byte-empty on clean / two named sections (## Boundary violations, ## Dispatch defects) when populated; non-zero exit on dispatch defects is the script-level fail-loud.
• G6 — stage-commit-parents validator: new scripts/validate-stage-commit-parents.sh with --capture --wave-id W<n> dual-write (key=value sidecar + OBC-shaped wave-<n>.txt) closes the Implement batch-gate bridge.
• G7 — anchor file lookup: SKILL.anchors.json schema + resolver harden the reference-gate pause + reference-extraction round-trip.
• G8 — version source unification: repo-root VERSION file is the single source of truth for .claude-plugin/marketplace.json, .claude-plugin/plugin.json, .github/plugin/*, and build/ outputs.
• G9 — active-skill prompt footprint <30,000 tokens (cl100k_base): 14 SKILLs trimmed via 4-pass (three-tier placement → script-mechanic deletion → R8 tightening → regression-guard execution) against six new _shared/ snippets !cat-resolved at skill-load. Heaviest active skill (using-qrspi) measured at 29,913 tokens.

Cross-goal decisions:

• CD-1 — vendor-neutral dispatch rename (scripts/dispatch-agent.sh, was scripts/review-prep.sh for review-only).
• CD-2 — implementer Pre-DONE self-check promoted from advisory to blocking.
• CD-3 — R8 prose-density rule landed in skills/_shared/prompt-design-rules.md with explicit ### What NOT to tighten carve-outs (load-bearing repetition, verbatim test-pinned strings, iron-law clauses, anchor phrases).

In-pipeline fixes captured (Integrate + Test rounds):

• FX-Integrate-F01 — SKILL printf %sn "$SHA" write shape for reviews/integrate/phase-base.txt aligned with OBC consumer (bare-SHA, no key=value).
• FX-Integrate-F02 — validate-stage-commit-parents.sh --capture --wave-id W1 + --seed-wave-1-obc mode dual-writes the OBC-shaped wave-1.txt bridge.
• FX-Integrate-CI — release pin updates (agents 42, workflows 2, marketplace 0.7.3); bats off-by-one CI failures (user-trap-EXIT vs bats EXIT handler collision in test-phase1-acceptance.bats [AC4]; heredoc-embedded @test count inflation in test-bats-test-name-id-hygiene.bats).
• FX-Test-FX1 — OBC report shape brought into compliance with design.md §G5(b) (byte-empty clean, two-section schema).
• FX-Test-FX2 — [Tnn] token violations in two unit fixtures runtime-assembled.
• FX-Test-FX3 — ### What NOT to tighten H3 subheading added to prompt-design-rules.md per CD-3.

Test posture: 89/89 plan-level acceptance tests GREEN; 2,566/2,566 full suite GREEN locally; CI BATS-under-bash-3.2 carries 9 pre-existing environmental baselines (root-uid permission tests + missing-tiktoken footprint tests) inherited from v0.7.2.

Reviews: Implement per-task reviews + 2-round Integration review + Test 3-reviewer fan-out + dual-review (Claude Opus 4.7 high + GPT-5.5) all clean.