v1.108.275 — A pattern that matches nothing now says so

#446. entry_point_patterns is documented as glob patterns and matched with fnmatch, which supports only *, ? and [seq]. Two constructs that work in every shell do not work here, and neither fails loudly: brace alternation is not expanded, and ** does not match zero directories (plugins/**/*.ts misses plugins/auth.ts).

A pattern that matches nothing is indistinguishable from a repo that genuinely has no such entry points — same output, same confidence, no marker. The caller gets more symbols reported unreachable and no reason to doubt it. We made the brace mistake ourselves, in shipped framework profiles, and did not notice for a release (#445).

Both dead-code tools now name the patterns that matched nothing, and the message explains both fnmatch surprises rather than only reporting the outcome. Reporting covers every cause at once — braces, **, a typo, the wrong path root — instead of the one spelling we happened to get wrong.

It reports; it never refuses. A pattern matching nothing is legitimate: a caller may pass one pattern set across several repositories.

Two gaps found while implementing it

get_dead_code_v2's existing message was correct and almost never fired. It already said "entry_point_patterns was supplied but matched no indexed file" — gated on entry_point_count == 0. Any repository carrying one ordinary main.py made the count non-zero, so the caller heard nothing. A correct warning behind the wrong gate reads as "no problem found". It is now ungated and names the offenders.

get_dead_code_v2's call-graph-only exit ignored the parameter entirely. When a repository has no import graph the tool returns early, through a path that never received entry_point_patterns and still does not use them — that mode has no file-level entry-point concept, so there is nothing for a path pattern to seed. That is defensible; silently accepting a parameter and discarding it is not. That exit now says so.

Both were found because the new end-to-end test landed on the fallback by accident. Every helper-level test passed while a whole exit ignored the feature, which is the argument for testing through the tool rather than the helper.

Not decided here

Whether entry_point_patterns should also accept brace expansion remains open on #446. Implementing the warning surfaced an argument for keeping it separate: a pattern containing literal {} matches such filenames today, so expanding braces would change existing behaviour rather than being purely additive — a real decision under the 1.x no-removal contract, and one that immediately raises whether ** should gain real recursive semantics too. The harm is closed either way.

Verification

tests/test_v1_108_275.py (14), 4 fail against the v1.108.274 call sites; the 10 passing on both sides are helper unit tests and controls, including one asserting the tools stay silent when every pattern matches.

Full suite 7740 passed / 8 skipped / 0 failed, and 7734 / 14 / 0 under a clean 3.13 environment (7748 collected both ways); ruff check src/ clean; all 9 CI jobs green.