Match weak-tier model-id markers as whole tokens (fixes Gemini misclassification)

[ishinder]

The problem

The brain exec-guard classifies a model as weak- or trusted-tier in
extensions/loom/exec-guard/model-tier.ts, and the weak-tier check matched its markers as
substrings of the model id:

if (WEAK_ID_MARKERS.some((m) => id.includes(m))) return "weak";

WEAK_ID_MARKERS includes "mini", and "gemini" contains it -- so
"gemini-3.1-pro-preview".includes("mini") is true and every Gemini model classifies as weak.
That check runs before the trusted-marker list and the price fallback, so it short-circuits: even
gemini-2.5-pro and gemini-1.5-pro, which are listed in TRUSTED_ID_MARKERS, come out weak, and
those two entries are unreachable dead code.

The user-visible effect: a weak-tier model never gets the trusted-workspace auto-allow (policy.ts),
so on any Gemini model "Trust this workspace" does nothing -- routine unknown commands
(curl ... | jq, or a script the agent just wrote) re-prompt on every call even in a trusted
workspace.

The fix

Match each weak marker as a whole, delimiter-separated token instead of a substring:

const idTokens = id.split(/[^a-z0-9]+/);
if (idTokens.some((tok) => WEAK_ID_MARKERS.includes(tok))) return "weak";

Every weak marker is a single token, so genuinely weak ids still match (gpt-4o-mini,
claude-3-5-haiku, gemini-2.0-flash, llama-3.1-8b), while "mini" no longer matches inside
"gemini". Gemini ids fall through to the trusted list and price fallback as intended, and the
gemini-2.5-pro / gemini-1.5-pro trusted markers are reachable again. The trusted-marker check is
left as a substring match -- those markers are longer, version-specific strings -- so there's no
behavior change for non-Gemini ids.

Tests

Extends tests/exec-guard-model-tier.test.ts with two assertions in the existing "frontier models
are trusted" block: gemini-2.5-pro and gemini-3.1-pro-preview must be trusted. Both fail on the
current substring check and pass with the token match (watched them go red, then green). Full suite
green (1237 passed / 1 skipped), npm run typecheck clean, prettier and eslint clean.

[dannon]

Good catch, thank you!