test: cover FallbackChainService behavior

[harshitethic]

Summary

• add coverage for FallbackChainService failure thresholds
• cover fallback progression and disabled entries
• cover success, reset, and clear state transitions
• cover summary and FallbackEntry display behavior
• isolate persisted UserDefaults state between tests

Closes #28

[macOS26]

Ran the suite locally on macOS: xcodebuild test -only-testing:AgentTests/FallbackChainServiceTests → 10/10 passed (✔ Test run with 10 tests in 1 suite passed).

I verified every assertion against Agent/Services/FallbackChainService.swift and they're all correct reads of the state machine: maxFailuresBeforeFallback = 2 (first failure → nil, second → enabledChain[0] with consecutiveFailures reset), the enabled-only filtering in recordFailure/activeFallback/summary, recordSuccess/reset returning to primary, clear wiping chain + state, and the "→ " active marker in summary. The FallbackEntry.displayName raw-provider fallback test is right too. The test hygiene is genuinely good: @Suite(.serialized), and the withIsolatedService snapshot/restore of both the singleton state and the raw UserDefaults keys is careful (the raw-data restore at the end makes the re-added-entries-get-new-UUIDs detail harmless).

One requested change — out-of-scope deletions. The diff removes ~25 lines of explanatory comments from existing, unrelated tests: the allCriteriaDone gate rationale in GoalStateStoreTests, the entire doc comment on ToolFailureDetectionTests explaining the twice-broken isToolFailure history (including why the bare word "failed" was deliberately dropped as a trigger), and the polling-exemption rationale in StuckGuardFingerprintTests. Those comments document regression history — they're the reason future contributors won't reintroduce those exact bugs. Please revert every change above the // MARK: - FallbackChainService line so the diff is purely additive (the one whitespace fix at line 111 is fine either way).

Also: the PR body says Closes #28, and PR #40 claims the same issue. Since FallbackChainService genuinely had zero coverage, this PR is the better fit for #28 — I've asked #40 to drop its claim, so keep yours.

Everything else is merge-ready; happy to approve as soon as the comment deletions are reverted.