QM-56: Hand the test-economy rules to every prompt that can write tests

[pcapriolo-yc]

Closes QM-56.

Changes

Why this matters: The factory's test-economy rules (fewer, better tests; the smallest test that proves the change) only reached the add-tests agent and the mutation-gate repair agent. But the agent that writes a run's code — implement on a normal pass, feedback-revise on a feedback re-run — usually writes the tests too, and it received no economy constraint at all. By the time the rules arrived, the bloat was already in the tree, and every later pass is additive rather than pruning. On QM-42 a feedback re-run produced 17 tests and 483 lines in a single file, including a 124-line fixture, before any agent had seen the rules.

What changes:

• The feedback-revise and implement agent prompts now carry TEST_QUALITY_RULES and TEST_HOWTO, so every agent that can write tests is held to the same standard before it writes anything.
• The review-kickback re-implement pass is covered by the same insertion, because it reuses the implement prompt and the rules sit outside the kickback branch.
• The rule text itself is unchanged and still defined in one place; no new constant, flag, or conditional was added.
• A guard in test/factory-layout.test.ts fails, naming the offending prompt, if feedback-revise, implement, or add-tests ever stops interpolating the rules.

Acceptance stories:

• On a feedback re-run, the Factory's feedback-revise agent used to get no test-economy rules; it now gets them before it writes its first test.
• On a normal pass and on a review kickback, the implement agent gets the same rules on both branches, where before it got none on either.
• The add-tests and mutation-gate repair agents keep the rules they already had, and the proof agent is untouched.
• A maintainer who removes the interpolation from any of the three prompts gets a failing test naming that prompt instead of a silently unenforced standard.

Test Plan

• NODE_ENV=test ALLOW_UNSIGNED_TEST_IDENTITY=1 node --experimental-test-module-mocks --test test/factory-layout.test.ts
• npm run typecheck && npm run lint
• Manual: removed each of the three interpolations in turn, confirmed the guard failed naming that exact prompt, and restored the tree.

Proof it works

The new guard passes on this branch and fails with the offending label named when either insertion is reverted, so the rules are provably delivered to both newly covered agents. This change alters only the prompt text handed to background agents, so it has no visible surface and produced no screenshots or recordings.