fix(ci): reserve the room a squash merge takes

[gloryfromca]

The gap

check_pr_title measures the title an author types. What lands on main is the subject GitHub builds from it on squash — <title> (#<number>) — which is longer by 6–9 characters. Nothing checks that one.

v1.4.2 went out through that gap. #496's title was fix(lancedb): cap both caches, give queries their own embed budget, a legal 66 characters. The commit on main is 73, over the 72 this same policy enforces for commit subjects — so commit messages then failed on every open PR, for a commit none of their authors wrote. #505 narrowed the range so those PRs stop being blamed; this closes the hole that produced the subject.

It is not a one-off

Of the 29 PRs open today, 5 would still land an over-length subject:

squash subject	PR
93	#445 fix(reflection): batch source deprecation, revert it on failure, cap sources per merge
91	#424 feat(memory): add Decision Memory Kind (extract → persist → index → search → evolve)
80	#470 build(deps): bump the python-deps group across 1 directory with 7 updates
77	#444 feat(llm): add aimlapi.com as an LLM endpoint with partner attribution
76	#419 build(deps): bump the actions group across 1 directory with 2 updates

Each one that merges puts commit messages back in the state #505 just cleaned up, for every PR behind it.

The change

Reserve the suffix before comparing. The workflow now passes PR_NUMBER, so the reserve is the real string — a four-digit PR costs one more character than a three-digit one. A local make check-pr-title has no number and falls back to the width of " (#0000)".

The message says what to do:

PR title is 66 chars; max is 65: fix(lancedb): cap both caches, give queries their own embed budget
  a squash merge appends " (#496)", and the commit subject it produces may not exceed 72

Verification

Five tests in tests/unit/test_scripts/test_check_pr_title.py, each failing against the current script:

• the fix(lancedb): cap both caches, give queries their own embed budget #496 title with its real number is blocked, and the message names the suffix
• a title that still fits once the suffix is counted passes
• the same title passes at #496 and fails at #4096 — the reserve is the real suffix, not a guess
• main() reads PR_NUMBER
• the pre-existing length test now expects the fallback budget

make lint and make check-commits pass.

Expect five PRs to go red

That is the point — they are red now instead of after they have put a bad subject on main. Each one needs its title shortened, which is a text edit on the PR, no code change and no rebase.