Fix module-level state leaks that broke randomized test order

[TheGreatAxios]

Summary

bun test --randomize failed with over a hundred failures spread
across unrelated files (measured at seed 42 on unmodified main:
~107-119 failures across repeated runs — the count is not perfectly
stable run-to-run, meaning some failures are timing-sensitive rather
than purely a function of the shuffle order). All of it traced to
two shared-state bugs, plus two further instances of the first one:

• mock.module restores that captured "the real module" via a live
  namespace binding — either await import() or a static
  import * as ns — were no-ops, because Bun mutates that same
  namespace object in place when the module gets mocked. The
  captured reference silently turns into the mock too, so a later
  mock.module(path, () => capturedReference) "restore" changes
  nothing. This broke:
  • tests/unit/tui/agent-tools.test.ts (authz/verify/secret-guard/
    read-file-guard plugins collapsed to fakes for every file that
    ran afterward)
  • src/tools/web-search.test.ts (dropped unwrapToolContent and
    connectMCPServers from mcp/client.js process-wide)
  • tests/unit/config.test.ts (node:os capture via a static
    import * as nodeOs, same defect, different capture shape)
  • mouse-reporting-disabled.test.ts (latent, not yet triggering a
    failure — fixed proactively)
  • src/session/state.test.ts mocked node:fs/promises with no
    restore at all; harmless in practice because its delegate was
    captured before the mock took effect, but still fixed to restore
    Fixed by shallow-copying the capture immediately ({ ...ns }),
    regardless of which import form produced it, and spreading the
    real module into the mock instead of replacing it outright. Also
    removed a dead mock of a src/web/plugin.js module that no longer
    exists.
• schedulePricingMetadataRefresh's one-shot latch is module-level
  and shared by the whole test process; any file that exercises the
  real loadConfig bootstrap path leaves it set, so
  pricing-metadata.test.ts's own dedupe test silently no-op'd when
  it ran after one of those files. Fixed with a beforeEach reset
  alongside the existing afterEach one.

AGENTS.md documents the rule in terms of any live namespace
binding, not just the dynamic-import idiom, so the next author
following the guidance doesn't reproduce the config.test.ts bug in a
different shape. CI keeps the deterministic fixed-seed step and adds
a nightly job with a fresh random seed each run, printed up front so
a failure reproduces locally with the same --seed.

No assertions were weakened. Each fix restores real isolation — the
tests that exercise the fixed plugins/latch still fail if that
isolation regresses.

Verification

• bun run typecheck, bun run build, bun run test all pass
• bun test ./src ./tests ./evals --randomize passes across 20
  seeds (42, 1337, 1361063567, 1-10, 100, 200, 300, 7, 555, 909090,
  2026, 314159, 8675309, 424242)
• No tests quarantined

Closes CL-5579

[linear-code]

CL-5579