RAI-303 (draft, do not merge): voice rules in the prompts and the style eval

Stacked on #19. Linear: RAI-303. DRAFT: it does not meet its own bar, so it is not merged.

What is here

• A "How you write" rule 9 in CORE_RULES, plus voice lines in the Collector, Mimi chat (now named for the distributor), Order and Credit skills.
• Delivery thanks template: "Thanks for confirming. Delivery X is now marked as received." flag_text was factored out.
• mimi/api/style_eval.py and eval/style/: 20 dev cases, 70 held-out cases (46, then 12 fresh before round 2 and 12 more fresh before round 3) and 8 hand-labelled calibration replies. Each reply is checked in code (dashes, banned phrases, figures unchanged, figures grounded, length, exclamation marks, no Pidgin) and by a Muse Spark judge (natural, professional, clear, plain English, on point). A case passes only if all of these pass.

Results (held-out; tuned on dev only)

• Round 1: dev 75%, held-out 74% (34/46). One invented figure, an eval fixture bug, now fixed.
• Round 2: dev 90%, held-out 72% (42/58); the 12 fresh cases alone 67%. One banned phrase ("Glad to help"); the guard was fixed for it.
• Round 3: dev 85%, held-out 74% (52/70); the 12 newest cases alone 67%.
• Dashes, figure changes and invented figures were zero in rounds 2 and 3. The judge agreed with all 8 calibration labels every round.
• On 7 held-out cases run on the old prompts with no guard: 3 to 4 of 7 passed, and 2 to 3 had em dashes.

Why it is below 90% Most misses are "on point" (an answer that leaves out a fact, such as the invoice number or the total next to the overdue figure) and "natural" (stiff openings, a bot-like note to the owner). A round-3 rule ("a polite ask") made things worse in places: it asked for payment on an account-details request. The judge is also strict on some code templates ("Delivery X is confirmed" counts as "system-written").

Not done The harness, context, credit and e2e caches were NOT re-recorded for these prompts (about N8,700), so CI evals on this branch miss their caches. Re-record only after the prompts pass. One held-out case (h3-col-pending-ref) was edited after recording: a real bank short name in its reference was replaced to satisfy test_no_real_banks, so it now misses its cache.