fix(arch): tighten test-file dirs and import-hint resolution

[Asgarrrr]

Summary

Three correctness fixes in senseArchitecture (V0.2 architecture sensing). All stay inside ADR 0005's filesystem-only envelope — no AST, no import graph, no new dependencies.

What changed

#	Heuristic	Before	After
1	test-file	Only ${dir}/<stem>... and ${dir}/tests/<stem>...	Adds co-located __tests__/ and repo-root tests/ / __tests__/
2	import-hint	basename(spec) === basename(changed)	Resolve spec against importer dir (POSIX), match resolved path with .ts/.tsx and /index.ts(x) suffix recovery
3	import-hint	Any from "..." regex match	Skip bare specifiers (from "react") and tsconfig-paths aliases (from "@/x")

Why now

A deep read of src/architecture/sense.ts surfaced two real correctness gaps and one real coverage gap:

• The previous ${dir}/tests/-only rule never fires on Mira's own repo — all tests live at top-level tests/. Same for any project using Bun-style top-level tests, Jest's __tests__/, or Python-style top-level tests/.
• Basename-only matching cross-reports collisions: changing src/utils/foo.ts flagged any file importing an unrelated src/auth/foo. Path-resolved comparison eliminates this without leaving filesystem-only scope.
• A changed file named react.ts matched from "react" purely by basename. Same fix.

tsconfig paths aliases are intentionally skipped — false negatives are preferred over false positives per ADR 0005's invariant. Documented as a deliberate non-coverage in a test case and in the ADR amendment.

ADR

Amended docs/adr/0005-architecture-sensing-v0-2-scope.md (V0.2.1):

• Bullet 3 (test-file): expanded directory list.
• Bullet 4 (import-hint): replaced basename matching with resolved-path matching.
• Added an Amendments section recording the change.

Test plan

• bun test tests/architecture-signal.test.ts — 27 pass / 0 fail (20 existing + 7 new)
• bun test — 316 pass / 0 fail / 23 files
• All existing import-hint and test-file tests pass unchanged under the new resolver — verified by tracing each case
• New cases prove F1 / F2 / F5 are fixed: they fail before the patch, pass after

Out of scope (deferred follow-ups)

• Pre-indexing source-file content once instead of re-reading per changed file (perf — O(n×m) → O(n+m)). Not a correctness issue.
• Side-effect (import "./x") and require() recognition. Would expand ADR scope.
• stemOf collapsing on first dot. Documented design choice.

Risk / rollback

The new import-hint matcher is stricter — projects that relied on basename-only matching for tsconfig-aliased imports will lose those signals. The ADR amendment makes this explicit; the V0.2 invariant "prefer false negatives" justifies the trade-off. Single commit, clean git revert.