check_docs.py reads the Step 4 substitution table through a fixed 8000-byte window with only 231 bytes of headroom #80

What went wrong

scripts/check_docs.py:63 scans for substitution-table rows inside a hardcoded byte window:

table_section = text[table_start:table_start + 8000]

Measured against create-dev-loop.md at c1ac200:

Step 4 section start: 42479
Step 5 section start: 50248
Step 4 section length: 7769 bytes
Window:               8000 bytes
Headroom:              231 bytes

The window currently overshoots the end of Step 4 by 231 bytes (harmlessly — it lands in Step 5's prose,
which has no table rows). But the GENERATED_AT row — the last real row — sits at offset 6558, leaving
roughly 1.2 KB before added rows start falling outside the window.

Why this matters

When a row falls outside the window, declared silently loses it and check_placeholders reports the
placeholder as missing a substitution-table row even though the row is right there. CI fails with a
message that points the contributor at the wrong problem — "add a table row" for a row that already exists —
which is a worse failure mode than not checking at all.

This is precisely the invariant CLAUDE.md calls load-bearing ("Every {{placeholder}} in the template must
have a corresponding row in the Step 4 substitution table"), so the check silently mis-firing as the table
grows is a real risk. Adding 3–4 more placeholders would get there.

Expected vs. actual

Expected: the Step 4 table is parsed by its actual boundaries.
Actual: it is parsed by a byte count that happens to fit today.

Suggested fix

Bound the section by the next ### heading rather than a byte count:

table_start = text.find("### 4 — Fill in the placeholders")
...
next_heading = text.find("\n### ", table_start + 1)
table_section = text[table_start:next_heading if next_heading != -1 else len(text)]

That also removes the current 231-byte bleed into Step 5, which is how the unused-looking
`path/to/file` example row (line 603, inside Step 4's DOC_CHECK_TABLE guidance) ends up counted as a
declared placeholder.

Worth a regression check that fails before the fix: append a dummy row past the 8000-byte mark and confirm
check_docs.py reports it missing, then confirm it passes with the fix applied.

Notes

Found during a Stage A documentation-accuracy sweep on 2026-07-29. Not fixed in that sweep's PR because it is
a code change, not doc drift.