chore: rewrite Tester to write and run automated tests

[wpinrui]

Why

Tricia guided the owner through long manual checklists. Her last plan for PR #62 ran to ~20 items and required a specific stat build (10 Charisma, or trait badges may never appear), a page-refresh trick to escape a negotiation without destroying the listing, and a warning that one check might never become reachable at all. That is not a viable use of the owner's time — and the tricks and warnings were symptoms of testability defects being handed to the owner instead of reported.

What changed

Automation is the default. Every check is triaged Automated or Manual. Manual requires a stated reason, and the only valid reasons are spatial layout, colour/legibility, responsive reflow, clipping, animation/audio, and feel. "Awkward to automate" is explicitly not a reason — it now routes to the Implementer as a testability defect.

The manual pass is a residual, capped at roughly eight items. Each item needs exact navigation ("Map → Scrapyard → Browse Junkers → click the first car", not "open a listing"), exact setup via the debug panel rather than grinding for a state, and a stated pass condition. Checks that need a workaround to perform are now bug reports, not owner instructions.

Feel questions are separated from the checklist and routed to the PM as product signal, rather than being mixed in as test items.

Tests may never be weakened, skipped, or deleted to get a green run. A failing test is the deliverable. Tricia still does not fix source.

Permission boundaries

Widened, scoped to testing:

• May write test files, test config, and test-only devDependencies (a DOM environment and a component-testing library, when a check genuinely needs rendered output — neither is installed today)
• May commit and push test files to the PR branch under test
• May not touch application source, app dependencies, or build config

Project notes recorded

• src/engine/ is pure and dependency-free; the RNG is seeded (Mulberry32) and injectable, so negotiation, pricing, and listing-lifecycle tests are fully deterministic with zero new dependencies
• createRNG() seeds from Date.now() and must never be called in a test
• Bare npm test starts vitest in watch mode and hangs — npm test -- --run is required
• No test files exist yet; Tricia sets the convention (colocated foo.test.ts)

Effect on PR #62

Against Tricia's own #62 checklist, the entire money path and the whole "one shot per car" rule collapse into unit tests. The owner is left with the four UI/layout items and one feel question.

Knock-on

Tricia now incrementally seeds backlog item 5 (the test/CI/data-validation foundation) as a side effect of every PR she tests. She does not replace it — there is still no CI and no runner config — but a suite will already exist by the time that ticket is picked up. Flagged to the PM in the handoff so Riley can frame the issue accordingly.