refactor: optimize goal/propose/tag — prose->table, dedup DRY_RUN, fix propose parallel-Skill claim

[JFK]

Problem

commands/goal.md (241), commands/propose.md (552), and commands/tag.md (425) carry ~1,450 tokens of compressible redundancy and a few precision gaps.

Scope (CTO/Explore review 2026-06-13)

• goal.md: the autonomy model is stated as a table then re-explained in prose (~goal.md:27-42); keep the table. Tighten the TDD-invariant paragraph (~goal.md:142-150) that restates the contract 3x. Make the gate1-vs-gate2 force-continue logic a table (~goal.md:180-182).
• propose.md — precision fix: remove the misleading "invoke both skills in parallel" instruction (~propose.md:288) — Skill is not parallel-capable in this harness. Clarify title-derivation (~propose.md:158) and the keyword/validation regex mismatch (~propose.md:124-126).
• tag.md: condense the repeated "skip if DRY_RUN -> bash -> verify" pattern (steps 7/9/10/11/12) and the long push-failure recovery section (~tag.md:324-338).
• Across all three: collapse prose+bullets pairs that restate the same parse logic into one pseudocode/table block.

Acceptance criteria

• Existing tests green; /propose, /goal, /tag behavior unchanged except the corrected parallel-Skill instruction.
• Token baseline (verdict parser: replace heuristic with structured ## Verdict: <token> contract #1) shows reduction.

Depends on

#1, #2.

[JFK]

Closing after re-audit (see #96). The two claimed "precision bugs" do not survive scrutiny:

• propose.md:288 "invoke both skills in parallel (two Skill tool calls in a single batch)" — this is correct, not misleading. Batched independent Skill calls are supported and were used in this milestone's own gate2 runs (cso/qa-lead/cto invoked in one batch). No fix needed.
• propose.md:124↔126 "regex mismatch" — not a contradiction. Extraction emits [a-z0-9 ] (lowercase alnum + space), which is a strict subset of the validation ^[a-zA-Z0-9 _.:-]{1,200}$, so extraction output always passes. The validation is loose, not wrong; the fallback is a harmless safety net.

The remaining items (goal.md autonomy table, TDD paragraph, tag.md DRY_RUN repetition, push-failure recovery) are either not actually redundant (the autonomy level bullets carry load-bearing forwarding detail) or load-bearing (push-failure recovery). The one genuinely-tabular item (goal.md force-continue) is shipped in #96. Reopen if a real precision issue is found.