fix(interview-mode): read attended/unattended as aliases for steering/auto

[MendixMau]

Direct lane

What changed and why:
CLAUDE.md, skills/conversion-runbook.md and intake Q9 told users to write Interview mode: unattended in PROJECT.md. But bin/interview-mode.sh accepted only steering | assist | auto and fell back to steering for anything else. A project that opted into unattended exactly as documented silently ran as steering.

What this PR changes:

• Aliases in every source. attended reads as steering and unattended as auto in $CLAUDE_INTERVIEW_MODE, .claude/.interview-mode, the PROJECT.md line, and --set (which stores the canonical name). Each alias prints one stderr note, e.g. interview-mode: 'unattended' read as 'auto' (alias).
• How the PROJECT.md line is read. Markdown bold and backticks are stripped and the value is lower-cased. Then:
  • A value that is exactly one mode or alias (a trailing . is allowed) is used, so **auto**, `auto`, auto. and Unattended all give auto.
  • Trailing text after steering or attended is a harmless note, because steering is the loudest mode.
  • Trailing text after assist, auto or unattended falls back to steering with one stderr line: "a quieter mode must stand alone on the line — write 'Interview mode: auto' and put notes on the next line". --explain shows fallback (quieter mode with trailing text).
  • An annotated or hedged line therefore never resolves quieter than it did on master. Typos and empty values still give steering.
• Machine-written sources stay exact: env, session file and --set.
• Output stays canonical, so the one consumer, claude-hooks/hooks/interview-gate.sh (which compares to auto), is unchanged.
• Docs. skills/interview-protocol.md holds the vocabulary and the "note may follow steering; assist/auto must stand alone" clause. CLAUDE.md and the runbook name Interview mode: auto (alias unattended) and point there. The intake Q9 body offers steering/assist/auto, with steering as the default.
• Q9 heading kept on purpose. It is sync's ## 9. locator and its question-text key.
• sync-project.sh and old Q9 bodies. Sync now treats any Q9 body that still opens with _Not yet asked._ as unanswered. Without this, the body rewrite would turn every untouched scaffold into the false "edited, NOT touching it" warning.
• bin/exec-approval.sh: comment-only change. It no longer claims to follow interview-mode's convention, and its parsing is unchanged.
• gate-check.sh:699 is deliberately unchanged. It is the byte-exact stale-scaffold reference.

README.md is not touched here. Open draft PR #334 is rewriting it, and #334's README already documents the canonical values.

Field evidence: none yet. This is a vocabulary mismatch found by reading during the 1.0 README review.

• No client data anywhere in the diff (leak guard clean)

• Size cap: 9 files, +223/−27 including the fixture and CHANGELOG

• Test tier: fixture level. tests/wave2/test-interview-mode.sh grew from 19 to 76 checks. They cover:

  • aliases in each source and --set storing the canonical name;
  • no alias note on canonical values, and the hook not blocking on unattended;
  • stand-alone bold, backticked and full-stop values;
  • steering plus a note, with no stderr output;
  • every hedge from both reviews giving steering;
  • the stand-alone message, on exactly one line, and the --explain source;
  • typos, auto-ish and empty values;
  • machine-written sources rejecting trailing text.

  Fixture	Result
  test-interview-mode.sh on master	59 pass / 17 fail
  test-interview-mode.sh on d3a75e1	62 / 14 (every leak case fails)
  test-interview-mode.sh on this head	76 / 0
  test-bug12-sync.sh	61 / 0
  test-exec-approval.sh	43 / 0
  test-interview-gate.sh	11 / 0

  render-routing.sh --check passes, with the baseline at 74,995 → 75,038 words (budget 80,000). check-scripts, check-portability and check-pr-discipline are clean.

• Instrument change: it adds no new parser of tool output. Paths are unchanged, and the code is bash 3.2 safe.

• CHANGELOG.md line appended, crediting MendixMau (found in the 1.0 README review)

Review

1. Independent review of 12838b4: MERGE-AFTER-FIX. The first-word read let hedged lines resolve quieter than on master, which breaks the header rule that a typo never resolves quieter. Examples: auto? (to confirm with user), unattended? ask user, auto | steering, auto (proposed, not confirmed). d3a75e1 answered with a hedge word list plus ?/| checks.
2. Confirmation review of d3a75e1: the word list leaked. auto / steering, auto or steering, auto (pending sign-off), (maybe), (tentative), (draft) and — awaiting user still gave auto. A list of hedge words cannot converge, because there are always more ways to hedge. 09691fc replaces the list with a rule that cannot leak: a quieter mode is only taken when it stands alone on the line. Trailing text is tolerated only after steering, where a misread costs nothing because steering is already the loudest mode.
   • Deliberate behaviour change versus 12838b4: plain notes after auto (auto (user asked 2026-10-14), **auto** — fast POC, Unattended, per user) now also give steering, with the stand-alone message. Telling a plain note from a hedge is exactly the problem the word list could not solve.
   • The fixture records every case from both reviews.

🤖 Generated with Claude Code

https://claude.ai/code/session_01VJgWP5vEoAsNsJYqCDMGNw