[MendixMau]

Fixes #676

What changed

• bin/gate-check.sh: the five Decisions-table readers all ended the table at any non-pipe line, blank lines included: has_confirmed_decision, has_auto_assumed_decision, auto_hand_pending, stage_exit_recorded, plus a fifth copy in check_stage_7 that the issue did not list. A CONFIRMED row after a blank line was therefore ignored. Stage 3/4 FAILed with "no Stage-N CONFIRMED decision", and Stage 7 read PENDING.
• The five readers now share one awk prefix, DECISIONS_TABLE_AWK, and read the register in two passes ("$f" "$f"). The first pass records which lines are separator rows, so the second can look one line ahead.
  • Blank lines are skipped.
  • A heading or prose line still ends the table.
  • So does a new header row (one followed by a separator) whose first cell is not Stage. An Open-questions table after a blank line is therefore still never read as Decisions.
  • auto_hand_pending now uses FNR for its line ids, because NR counts both passes.
• New fixture tests/wave2/test-decisions-blank-line.sh:
  • D1: a blank line, then a Stage-4 CONFIRMED row. Expected: Stage 4 PASS.
  • D2: the same for Stage 7.
  • D3: a blank line, then an Open-questions table with | 4 |…CONFIRMED| and | 7 |…CONFIRMED|. Expected: Stages 4 and 7 do NOT pass.
  • D4: an auto-mode ✋ row after a blank line. Expected: it is listed at PROJECT.md:6.

Evidence

• Positive control: against origin/master ee12a61, D1, D2 and D4 FAIL and D3 passes. D3 is the guard: against a blanks-only variant of the fix, D3's two assertions FAIL, because the Open-questions rows then pass Stages 4 and 7.
• Fixtures on this branch:
  • test-decisions-blank-line: 5 ok
  • test-bug02-register: 8 ok
  • test-bug03-gates: 99 ok
  • test-auto-mode-hand-gates: 28 passed
  • test-wrong-verdicts: 25 passed
  • test-pipefail-grep-q (bin/lib/skill-routing.sh): 3 passed
  • check-scripts: 129/129 parse
• Leak guard is clean, and check-pr-discipline is clean.
• I reproduced the bug the same way the issue reports it, on a scratch register: Stage 4 FAIL / Stage 7 PENDING → PASS / PASS. I did not run it on a real project.

Time cost (#687)

• No new check, gate or obligation. This only changes how existing table reads parse.
• Runs once per gate-check invocation (each reader reads PROJECT.md twice instead of once). Measured on the same scratch project: gate-check.sh p1 4 averaged 921 ms before and 844 ms after, so any difference is below the noise.
• It cannot add a red verdict for a correctly written register. It removes a false FAIL, and with it the gate fight of a CONFIRMED row in plain sight.

Regression risk after sync

• Verdicts that flip: a register whose Decisions table is split by blank lines now has the later rows counted. Some Stage 3/4 FAIL lines become PASS, a Stage 7 PENDING can become PASS, and stage_exit_recorded can now raise the "approved over a failing gate" flag for such rows. The auto-mode ✋ list (and with it the Stage 6 FAIL) can list rows it used to miss.
• No new FAIL for content that was right, apart from the auto-mode ✋ rows above, which were already supposed to be listed.
• Unchanged:
  • A register with no | Stage | header falls back to scanning every table, as before.
  • Header rows without "Stage" are skipped, and they never hold CONFIRMED.
  • Exit codes and the output format are the same. Line ids from auto_hand_pending are unchanged.
• Portability: POSIX awk only (FNR == NR, array lookahead), no gawk-isms. It does not depend on the project layout: the register comes from the existing REGISTER resolver.

Review: pending independent review

---

Generated by Claude Code