fix: inline regex flags ((?i)) silently killed 50 of 57 instincts

[juanparisma]

The bug

(?i) is Python/PCRE syntax. JavaScript has no inline flags:

new RegExp("(?i)foo", "i")   // SyntaxError: Invalid regular expression: Invalid group

Both activators compile trigger patterns inside a try/catch whose handler is a
bare continue:

// core/_instinct-activator.sh:118
if (!new RegExp(tp, "i").test(context)) continue;
} catch(e) { continue; }          // ← the instinct disappears, and nothing says so

So an instinct written with (?i) is dropped in total silence. It stays in the
index looking perfectly healthy — valid JSON, right level, sensible inject — and
never fires once for the rest of its life.

Why it went unnoticed

Nothing surfaced it. The index was valid, /dream reported only the generic
invalid_regex, and the occurrence counter stayed at 0 — which is exactly what an
instinct whose trigger simply never came up also looks like.

Trigger patterns are authored by the model when an instinct is created, and (?i) is
the reflex spelling for anyone used to grep or Python. So it accumulates quietly, one
instinct at a time.

On the index that surfaced this, 50 of 57 instincts were dead. Sorting by
occurrences made it unmistakable:

occurrences	pattern	instinct
2213	no (?i)	env-vars-never-hardcode
642	no (?i)	error-handling-explicit
552	no (?i)	git-push-rebase-shared-repo
548	no (?i)	git-commit-conventional
390	no (?i)	memory-md-index-only
249	no (?i)	verify-numbers-at-source
0	has (?i)	…the other fifty

The survivors were simply the ones nobody had prefixed.

There is a second-order hazard worth flagging: because everything looked
never_activated, /dream's staleness module marked ~50 instincts as
archive_candidate. Following that advice would have archived a knowledge base whose
only problem was a syntax error.

The fixes

• Both activators normalise inline flag groups before compiling. Stripping (?i)
  is lossless here — they already pass the "i" flag, so it was redundant on top of
  being fatal. Other groups map to real flags where JS has an equivalent and are
  dropped otherwise, which still beats losing the whole instinct. This revives
  existing indexes in place, without asking anyone to edit their data.
• A pattern that genuinely cannot compile is logged as BAD_PATTERN with the
  instinct id and the engine's message. Silence is what let this run for months.
• /dream distinguishes inline_flag_group from invalid_regex and carries a
  Detail column, so the report names the cause instead of a category.
• commands/analyze-session.md tells the author not to write inline flags, so new
  instincts stop arriving broken.

_passive-activator.sh gets the same treatment. The shipped rules are clean today,
but they are hand-written like the instincts — the next one is one (?i) away from
vanishing the same way.

Verification

tests/test-inline-regex-flags.sh pins all of it: that JS really does reject the
syntax, that a (?i) instinct fires, that an uncompilable pattern is logged, and that
/dream separates the two causes.

• 4 of 5 fail against the unpatched activators, 5 of 5 pass with them — checked by
  stashing the fix, so the tests are not passing for their own reasons.
• All 17 suites green locally on a clean clone at a038b50 (16 existing + the new
  one, which is wired into .github/workflows/tests.yml).

Scope

6 files. No data migration: users' indexes keep their existing patterns and start
working again on the next tool use.