perf: skip rules below the minimum visible severity

[mattiamanzati]

What changed

In tsc CLI mode without includeSuggestionsInTsc, suggestion- and message-severity diagnostics are dropped from the output — but only after their rules had already executed, wasting their type-checker queries on every file. This PR skips those rules before execution:

• rulerunner.Run now takes a minSeverity argument: the least visible severity the caller can surface, computed via the new rulerunner.MinVisibleSeverity(config) helper (CLI without includeSuggestionsInTsc → warning; LSP/tests/oxlint and CLI with suggestions enabled → message, i.e. nothing is skipped).
• A rule whose config-resolved severity is below the threshold is skipped unless any directive in the file references it (by name or *, at file/section/next-line level, regardless of assigned severity). This guard is needed for two reasons: a directive like // @effect-diagnostics ruleName:error can raise the rule above the threshold, and a suppression directive must still be marked used so unusedDirective reporting stays accurate (skipping the rule would fabricate a spurious "directive has no effect" warning).
• transformDiagnostics now filters by the same threshold instead of consulting global CLI-mode state, so the runner is mode-agnostic and the previously hard-to-test CLI behavior is covered by plain unit tests.
• New Severity.AtLeastAsVisibleAs in etscore with an explicit visibility ranking (error > warning > suggestion > message > off/skip-file), since the constants' declaration order is not a visibility ordering.
• The existing skipDisabledOptimization escape hatch also disables this skip, matching the off-severity skip.

Example

// suggestion-severity rules are skipped entirely during `tsc` builds…
const x = Effect.succeed(1)

// …unless a directive references them, e.g. raising one to error:
// @effect-diagnostics someRule:error
const y = Effect.succeed(2)   // someRule runs here and reports as error

Verification

• Emitted diagnostics are byte-identical (errors and warnings) to a pre-change baseline across a full tsc -b --diagnostics build of the Effect monorepo.
• Measured on that build (3-run medians): ~1–2s less check time and −23.5M allocations vs the same-day baseline (this box had ambient load, so the wall-clock delta is a range; the allocation drop is load-independent evidence the skip is active).
• New unit tests: severity ranking, directive-mention matching (HasAnyDirectiveForRule), and MinVisibleSeverity under CLI mode toggled via etscore.EnterCommandLineMode — the CLI-mode test that previously blocked this optimization.
• pnpm lint, pnpm check, and the full pnpm test suite pass.

🤖 Generated with Claude Code