Add /flow-card: author a card's completion contract before triage

[dinershtein]

What failure mode this closes

A rough card is underspecified in exactly the ways triage detects and cannot fix alone, so it pays a three-command round-trip: /flow-check → [meta] QUESTIONS → Human Questions → /flow-questions → Backlog → /flow-check again. The questions are the same either way; the human just answers them hours later, without the context that made the intention obvious when they typed it.

/flow-card <rough intention> moves that Q&A to authoring time:

1. Ground the sentence in the repo (files must exist now; matching learnings.jsonl pitfalls are read).
2. Sort the eight card-eval.md gap categories through the same auto-answer policy triage uses — safe defaults become [meta] ASSUMED comments, only What / Why / Dependencies / split-implying Size reach the human.
3. At most three AskUserQuestion turns, each with a Park in Icebox escape.
4. Show the composed card in full and gate create_item behind a Create it / Revise / Discard approval.
5. Create one Backlog card carrying a completion contract: ## Goal, ## Why, ## Done when, ## Out of scope, ## Stop if, ## Context.

Deliberate limits

• Not a second planner. No approach, no file manifest, no test commands — the WHAT/HOW split and the human's plan-approval gate are untouched.
• Exactly one card per invocation. An intention worth several PRs is narrowed to its first shippable slice, the rest parked in Icebox. /flow-check's SPLIT path stays the single tested splitter.
• The three-turn cap can overflow. A card may still reach triage with an open question — it lands as a ## Stop if bullet and the existing QUESTIONS route handles it.
• Optional. A hand-written card behaves exactly as before.

No new config keys (config-contract.txt untouched), no new states, no changes to any existing prompt. create_item / update_title are already required provider ops, so Trello and GitHub both work; native_ids: yes suppresses the [PREFIX-N] title prefix automatically.

Prior art

The contract shape (end state / proof / boundaries / loop / stop rule), "define proof, not effort", show-the-exact-text-before-starting, and the "a prose menu of options is a defect — use AskUserQuestion" rule are adapted from MoonshotAI/kimi-code's builtin write-goal skill (MIT). Re-implemented against this kit's card/PLAN architecture — nothing copied verbatim, and goal mode's runtime half has no analogue here (agent-flow's driver is /flow-run).

Verification

bash tests/kit-invariants.sh   # 154/154
bash tests/kit-bin.test.sh     #  49/49

kit/commands/flow-card.md is 236 lines against a new 270-line budget in tests/kit-invariants.sh.

Contract lines that moved

None. This PR adds a command and its budget; card-eval.md, plan-format.md, flow-check.md, flow-run.md and every worker prompt are unchanged. docs/design-notes.md records problem → mechanism → limit.

🤖 Generated with Claude Code

https://claude.ai/code/session_01Tdu4p6jdzNSQd7P71iBw8G

[dinershtein]

Self-review before asking for a second pair of eyes — two defects found and fixed in 443fe46:

1. The approval gate contradicted two exits added before it. Park in Icebox, a cancelled prompt, and the too-big branch all said "create in icebox" from step 5 — skipping compose and approval, and contradicting the contract line one screen above. All three now change the destination, not the route.
2. ## Done when asked for proof without saying where to find it. Step 2 now reads the target's test_cmd/lint_cmd from tracker.json and notes what the repo already uses to prove this kind of change.

Not covered by anything here: this command has never created a card. The kit's tier-2 (behavioral) checks are deliberately out of CI, so everything green above is static.