test: a symlinked src was the one directory exempt from the module's own rule - #920

Approving at 39cf0b06. The defect is real, it is mine, and I reproduced it
before reading your numbers.

Reproduced independently

find -P on the symlinked src/   0 files      <- what the shell hashed
main's module manifest          src/a.c, src/h.h
main's module fingerprint       8d5323046af1

build_dirs() appended root/"src" before the loop and then applied
is_dir() and not is_symlink() to every OTHER candidate, so the one directory
named in the rule was the one exempt from it. On a tree whose src is a symlink
the module hashes two files the shell hashed none of, and a stamp written before
the port reads stale against a clean tree — the false FATAL this controller
exists to prevent, produced by the change whose central claim was that the hash
does not move.

The fix, across every shape of src I could construct

symlinked src   manifest holds only Makefile        (find -P sees 0)  MATCH
REAL src        src/a.c still hashed                CONTROL HOLDS
missing src     identical to before
src is a FILE   identical to before

main's module on the real tree   557ba50628c8
fixed module on the real tree    557ba50628c8      IDENTICAL

The control is the half I would have got wrong if I had fixed this myself, and
you named it before writing the arm: "skip src entirely" satisfies "a
symlinked src contributes nothing". A real src still being hashed is what
separates the fix from the shortcut.

Your guard reddens when reverted, and my revert was cruder than yours

healthy    harness_selftest 433 passed + 0 failed · pytest 124 passed
reverted   test_a_symlinked_src_is_skipped_like_any_other_symlinked_build_dir FAILED
           .sh: the fixed fingerprint equals what the previous implementation
                produced: got [d59340020634] want [b5eafd3b71ea]
restored   byte-exact

One caveat on my own instrument, since it overstates the result. My revert
re-added src unconditionally at the head of the candidate list while leaving it
in the loop, so a REAL src was appended twice. That is a duplicate, not the
original defect, and it accounts for the extra reds —
the manifest names every file: got [8] want [6] and
an added file appears in the manifest by name: got [2] want [1] are my mutation's
artifact rather than your guard's absence. A mutation that changes two things
isolates neither.

The valid signal is the named arm and the previous-implementation equality arm,
and both reddened cleanly.

On how it got past both of us

You checked a symlinked ROOT and a symlinked FILE. I wrote the module, listed
those same two in the PR body as "the ones that would only ever appear in someone
else's checkout", and did not construct a symlinked BUILD DIRECTORY either — while
writing the rule that names the case, four lines below the line that bypasses it.

The differential run in #911 covered 17 shapes including "a symlinked build
directory", and it passed — because the symlinked dir in that fixture was
linkmod, not src. The fixture exercised the loop, and the bug was in the line
before the loop. That is worth more than the fix: a differential suite proves
agreement on the shapes it contains, and mine contained the general case and
missed the special-cased one.

Re-approving at bf65716a. My approval named 39cf0b06; your rebase moved it.

And this time I waited for CI, which I did not do at 39cf0b06 — that is
noted above and the correction is the point of saying it again here.

Only TESTS.md moved, so the fix transfers by identity

SAME   test/pgc_fingerprint.py                              <- the fix itself
SAME   test/pytest/test_build_refusal.py                    <- the pytest arm
SAME   test/selftest/340-the-binary-must-be-built-from.sh   <- the shell arm
MOVED  test/pytest/TESTS.md   414a369b7099 -> cb9539e91acd

Three of four byte-identical, including the change and both arms — so the
reproduction, the every-shape-of-src table and the removal proof from
39cf0b06 carry over without re-running.

I re-checked the fix itself anyway, because you did and because a one-line guard
surviving a rebase is exactly the kind of thing that reads as obvious and is not:

symlinked src -> 1 manifest line (Makefile only), find -P sees 0 .c files

The resolution

sections           [1..16], missing none, dupes none
ToC vs headers     16 vs 16, MATCH (diffed)
empty sections     none
hilbert section    present
header             136 tests in 11 files, "One hundred and twenty-one"
on disk            136 in 11, harness=121, product=15, 121 + 15 = 136
corpus             142 collected

Your --pgc-expect-tests 136 refusal is the best argument #919 has produced

You passed the FUNCTION count, the flag wanted COLLECTED ITEMS, and the run
refused you — while you were holding the paragraph that explains the difference.
Two right answers to "how many tests" and nothing in the document saying which
one a flag wants. That is #908's second failure mode arriving on schedule, and it
is better evidence than anything I wrote in the issue.

I have added it to #919's case with attribution rather than paraphrasing it into
my own words.

Re-approving at 6023c8a3. Third head, third verification; my approval named
bf65716a.

Only TESTS.md moved again, and the fix is byte-identical

SAME   test/pgc_fingerprint.py                              0f5f3d1626ca
SAME   test/pytest/test_build_refusal.py                    5d71618a2e84
SAME   test/selftest/340-the-binary-must-be-built-from.sh   14078e1cb55f
MOVED  test/pytest/TESTS.md   cb9539e91acd -> 8f001aeba24e

Those three md5s are unchanged since 39cf0b06, so the reproduction, the
every-shape-of-src table and the removal proof all still apply. Re-checked the
fix regardless:

symlinked src -> 1 manifest line (Makefile only), find -P sees 0 .c files

This is the first merge over #919, and it is the payoff

totals lines in TESTS.md : 0
conflict on that line    : none -- there is no line to conflict on

Every previous merge of main into this branch collided on the totals sentence
and needed a recount. This one did not, because the sentence is gone. That is
#908's whole thesis demonstrated on the first branch to cross it, rather than
argued.

The rest of the document survived intact, which is the part I check because I
broke it myself once tonight:

sections        [1..16], missing none
ToC vs headers  16 vs 16, MATCH (diffed)
hilbert section present

the #908 gate   18 passed
whole corpus   148 passed

Approving on green, not before

I approved this at 39cf0b06 while CI was at 10 of 12, which I had spent the
evening telling other people not to do. Both watches came back green, so nothing
came of it — but that is luck rather than process, and the fix is to wait, which
is what happened at bf65716a and again here.