feat(dev): port dev workflow to compiled spec→plan→run.js (v5.56.0)

[edwinhu]

Summary

Ports the dev workflow to the compiled spec → plan → run.js pattern — the second concrete instance after ds (#7). Replaces the generic-interpreter dev-implement.js (LLM "discovery" → per-level → heavyweight verifier) with a deterministic compiler + a per-plan compiled runner. Built parallel to ds (own scripts/dev/), not a premature shared-core extraction (that is the explicit next pass — ds + dev now triangulate the real seams).

What changed

New

• scripts/dev/dev_plan_table.py — tolerant parser, dev columns (Task|Deps|Files|Failing Test|Verify Command|Implements) + lifts ## Global Constraints + per-task ## Task Interfaces. 32/32.
• workflows/templates/dev-run-template.js + scripts/dev/dev_compile.py — sequential-within-level driver (shared tree, no worktree-merge), TDD test-first implementer that inherits the session model (no tier), independent probe gating on the real Verify Command exit code ({exit0, filesPresent, testPresent}), R4 / declared / hybrid-fullsuite pauses, stale-gate backstop, no-phantom-RED rule. Driver 35/35.
• docs/DESIGN-dev-spec-plan-compile.md — design + full parity log (§8c).

Changed

• hooks/dev-plan-executable-guard.py reconciled to import the shared parser (compiles ⇔ passes gate) — closes the LLM-discovery format-drift mask.
• skills/dev-implement/SKILL.md slimmed to COMPILE → run/pause; G6 doc-drift fixed (guard docstring + plan-template no longer claim worktree/parallel/merge).

Removed

• workflows/dev-implement.js — retired after parity proven; no live caller.

Why (wc-audit, baseline 7.5/10)

G1 honesty fix: the old gate keyed on the implementer's self-reported verifyPassed + a non-executing corroboration — no component ever ran the Verify Command into the gate. The independent probe actually runs it, making "real exit codes, not self-judgment" true for the first time.

Parity — PROVEN on the real hylo-tauri repo

• AB1 idempotent gate-first skip
• AB2 R4 stale-gate backstop — implementer honors a decision in code then re-blocks on a stale Verify (fails loud, never bends code); full edit→recompile→green round-trip
• AB3 payload carries deviations + test numbers (not a boolean)
• Hybrid full-suite cadence both ways — terminal overlap → suppress; non-terminal overlap → pause→(real suite)→resume→advance; skipped level correctly doesn't re-fire (levelDidWork=false)
• Bonus: a split-brain [object Object] bug that tsc+unit pass but integration breaks — live evidence for keeping the adversarial layer outside run.js

All ds safety invariants carried; ds suites unaffected (15/15, 28/28, grain-pause 12/12).

🤖 Generated with Claude Code

https://claude.ai/code/session_01UBQQ5zUqZYQjHuZaCrgbkJ