refactor: enforce canonical uniqueStrings helper and remove local copies

[thymikee]

Problem

uniqueStrings already has a canonical implementation in @agent-device/kernel/collections, but three production modules still carry equivalent local copies:

• packages/selectors/src/internal/build.ts
• src/platforms/apple/core/perf-frame.ts
• src/platforms/web/agent-browser-lifecycle.ts

The duplication survived a prose-only AGENTS.md instruction, demonstrating that contributor guidance is the wrong enforcement layer. All three owners can consume the kernel helper; @agent-device/selectors already depends on @agent-device/kernel.

Outcome

Make the canonical helper the only production implementation and enforce that boundary mechanically, so another local copy cannot silently appear.

Acceptance criteria

• Replace the three local implementations with @agent-device/kernel/collections imports without changing ordering or deduplication behavior.
• Remove the local helper declarations and any tests made redundant by the shared implementation.
• Add the smallest owning lint, fallow, or structural rule that rejects a production reimplementation of this exact helper pattern without reconstructing general semantic equivalence.
• Plant a representative duplicate and observe the owning check fail before restoring the fixed tree.
• Run pnpm check:affected --run and report the resulting gate plan.

This is a follow-up from PR #1888's agent-context review; it is intentionally separate because that PR changes contributor guidance and gate wiring, not unrelated production modules.

[thymikee]

Closing: the dedup landed in #1892 (d07281963), and the enforcement criterion is dropped rather than deferred.

All three local copies now import @agent-device/kernel/collections, with ordering and deduplication behavior unchanged (packages/selectors/src/internal/build.ts, src/platforms/apple/core/perf-frame.ts, src/platforms/web/agent-browser-lifecycle.ts). The one semantic difference is deliberate: the web module’s local copy also filtered empty strings, and that filter is now composed at the single call site that receives caller input.

On acceptance criteria 3 and 4 — the owning lint rule and its planted-duplicate red proof — a rule is not worth building for this:

• A lane whose whole job is to reject one two-line helper cannot state a Catches/Evidence/Cost/Kill-criterion case (Prune the test lanes that don't pay for themselves, and add the checks we're missing #1781). Its catch set is a single identifier.
• The recurrence evidence argues against it, not for it. The other two helpers under the same AGENTS.md rule (inferFillText, evaluateIsPredicate) have exactly one definition each; only the trivially re-typeable one drifted.
• The available mechanisms are all a poor fit. oxlint 1.69 implements neither no-restricted-syntax nor id-denylist; fallow rule packs only express banned-call / banned-import, not a local declaration; fallow duplication needs floors (--min-tokens 5 --min-lines 1) that are unusable repo-wide. That leaves an experimental jsPlugin lane — new tooling surface to guard fifteen tokens.

The premise that prose guidance is the wrong enforcement layer still holds in general; it just does not justify machinery at this size. If a third copy of this helper ever appears, that is the evidence to revisit — and it would be a new issue with a real recurrence count behind it.