fix(herdr): match bare composer prompt glyphs by alternation, not a bracket class

[zachlandes]

What Changed

• FM_BACKEND_HERDR_BARE_PROMPT_RE now defaults to the alternation '^(❯|›)' rather than the bracket class '^[❯›]'.
  Under a byte-oriented locale (LC_CTYPE=C, which the supervise daemon inherits) grep walks a bracket class byte-wise, so [❯›] matched any row whose first byte was 0xE2 — including the box-drawing glyphs ╭ ╰ │.
  Because fm_backend_herdr_composer_state lets the lowest matching row win, a bordered composer's own closing border overwrote the correct shape=bordered verdict and was carried in as content: an empty composer read pending (away-mode defers delivery), and a dark-themed composer holding real typed text read empty (away-mode injects over an unsubmitted draft).
  The alternation matches each full glyph sequence, so it is correct in any locale and matches a strict subset of the old pattern's matches.
• tests/fm-backend-herdr.test.sh adds three composer_state cases that pin LC_ALL=C explicitly, which no composer case did before.
  Two are the regression pins and cover both failure directions (empty bordered composer reads empty; a dark-themed composer with real text reads pending); the third pins that the alternation still matches each full unbordered ❯ / › prompt glyph, so a pattern matching nothing could not satisfy the other two.
• docs/herdr-backend.md records the 2026-07-16 incident and a known-gap note, and docs/configuration.md warns that an override of this knob must be spelled as an alternation.
  The gap: the composer's two grep call sites (fm_backend_herdr_composer_state and fm_composer_idle_matches in bin/fm-composer-lib.sh) still inherit the ambient locale, unlike fm_composer_strip_ansi and fm_composer_strip_ghost which pin LC_ALL=C.
  The shipped defaults are locale-invariant, so fleet behavior is correct, but an operator override of FM_BACKEND_HERDR_BARE_PROMPT_RE or FM_BACKEND_HERDR_IDLE_RE still carries locale-dependent semantics; pinning those call sites is deferred to a separate change against the shared lib.

Risk Assessment

✅ Low: The one-line fix is verified correct and matches a strict subset of the old pattern's matches (removing only false positives, so no previously-correct read can change), and the follow-up commit adds locale-pinned regression tests that I traced as genuinely failing on the old regex and passing on the new one, with scope held exactly to the captain's deferral decision.

Testing

Completed 1 recorded test check.

Pipeline

Updates from git push no-mistakes

⏭️ **intent** - skipped

✅ No issues found.

✅ **Rebase** - passed

✅ No issues found.

🔧 **Review** - 2 issues found → auto-fixed ✅

• ⚠️ bin/backends/herdr.sh:746 - The fix ships no regression test, and the existing suite cannot distinguish it from the bug. No herdr/composer test pins a locale, so everything runs under ambient UTF-8 where the old '^[❯›]' and new '^(❯|›)' behave identically — which is exactly how this regression shipped. The coverage is already written but runs in the wrong locale: the fixture at tests/fm-backend-herdr.test.sh:818 (' ╭───╮ / │ ❯ │ / ╰──── Composer ───╯') is the precise triggering shape, and under LC_ALL=C with /usr/bin/grep the old regex makes it return 'pending' where the test asserts 'empty'. Adding a composer_state case that runs under LC_ALL=C would fail on the old regex and pass on the new one. This also departs from the repo's own convention for this area: every prior composer fix (0eaf293, a955a05, 1811b89) shipped a regression test in tests/ plus a docs/herdr-backend.md entry.
• ℹ️ bin/backends/herdr.sh:839 - This grep is the only text-processing step in the composer pipeline that does not pin a locale — fm_composer_strip_ansi pins 'LC_ALL=C sed' (fm-composer-lib.sh:61) and fm_composer_strip_ghost pins 'LC_ALL=C awk' (fm-composer-lib.sh:87), both with comments explaining why. Inheriting the ambient locale here is what made this bug a heisenbug: UTF-8 in an interactive shell (works), C under the daemon (broken). The alternation makes the default pattern locale-invariant, but FM_BACKEND_HERDR_BARE_PROMPT_RE is an advertised override knob, so a user-supplied pattern still has locale-dependent semantics; the same applies to FM_BACKEND_HERDR_IDLE_RE via the unpinned grep in fm_composer_idle_matches (fm-composer-lib.sh:177). Pinning LC_ALL=C at the call site would make override behavior deterministic and fail visibly in dev rather than only in the daemon. Noting rather than pushing: the author's comment ('correct in any locale') plus the docs warning is a defensible alternative design.

🔧 Fix: add C-locale composer regression tests and incident entry
✅ Re-checked - no issues remain.

✅ **Test** - passed

✅ No issues found.

• command -v tmux >/dev/null || { echo "tmux is required for e2e tests" >&2; exit 1; }; tmux -V; rc=0; for t in tests/*.test.sh; do echo "== $t =="; bash "$t" || rc=1; done; exit "$rc"

✅ **Document** - passed

✅ No issues found.

✅ **Lint** - passed

✅ No issues found.

✅ **Push** - passed

✅ No issues found.