fix: duel domain plumbing + manual fame_score overrides

[Calixteair]

What

Two related fixes after the phase 3 multi-domain ship.

1. Duel page now respects the duel's actual domain instead of hardcoding the page default. A RER duel link opened from a paris-metro page used to mount Grid in paris-metro mode and clobber the wrong localStorage key. Also adds a domain picker on /duel landing so the user can intentionally create RER duels.
2. Manual fame_score overrides for the outliers the auto-ingest got wrong (Saint-Sulpice eclipsed by the église, Auber / CDG / Antony dropped by the topic filter, etc.). Curator decisions survive future re-ingests thanks to a new domains/<id>/fame_overrides.json convention.

Why

These are the two real-world bugs that surfaced once the phase 3 multi-domain UI hit prod. The duel one was structural (any future RER duel was broken); the fame one was cosmetic but visible — Saint-Sulpice at fame=1 on a station obviously well-known.

How

ceafeae fix(web): duel uses duel.domain, not the page default

`DuelView.svelte` now tracks two state vars:

• `viewDomain` ← `GET /api/duels/{id}.domain`. Drives the playing screen.
• `createDomain` ← page prop initially, then mutated by a new fieldset picker on the landing screen.

The fieldset (`kd-domain-picker`) renders one pill per DOMAINS entry above the Create CTA. Hidden when only one domain is configured. Active pill = primary background.

`` is now mounted with `domain={viewDomain}` — never `{domain}` (the page prop). i18n: `duel_create_pick_domain` key fr + en.

ESLint's `svelte/valid-compile` flags both prop captures as initial-only references; that's intentional and disabled inline with a comment.

`7bf409f feat(domains): manual fame_score overrides`

New file `domains//fame_overrides.json` shape:

```json
{
"_doc": "...",
"_rationale": { "": "explanation" },
"overrides": { "": 80 }
}
```

Loaded by both ingest scripts (`fame_score_paris_metro.py`, `fame_score_rer.py`) after the percentile-rank pass:

• Validates each value is `int in [0,100]`
• Merges into `fame_by_id` before patching entities
• Annotates the audit report with an `"override": true` flag

8 entries shipped:

• paris-metro: `saint-sulpice` → 80 (was 1, eclipsed by the church article)
• rer: `auber` 92, `aeroport-charles-de-gaulle-1` 95, `antony` 85, `cergy-prefecture` 75, `issy` 70, `issy-val-de-seine` 60, `saint-michel-notre-dame` 95

The `entities.json` files are patched in this commit too so the change is visible in prod without waiting for the next manual ingest run.

Checklist

• Conventional commit titles, no AI / Claude mention
• `cargo fmt --check && clippy --all-targets -- -D warnings && test --workspace` passes (109 tests)
• `pnpm --dir web lint && typecheck && test && build` passes (25 tests, 17 pages)
• No new `unsafe`, no `unwrap()` in non-test code
• No new secret committed
• No contract change (`duel.domain` was already declared in OpenAPI)
• Mobile-first respected (picker pills wrap to a 2nd row on 360 px)

Test plan

• Create a duel on `/duel` → pick "RER d'Île-de-France" pill → CTA → share link points at /duel?id=…&sig=… → opening the link plays a RER grid
• Open the link in another browser → 'Play this grid' starts a RER game, localStorage key is `kalidoku.game.rer.v1`
• Open the same paris-metro daily on /paris-metro/ → solve Saint-Sulpice → fame chip says 80 (was 1)
• Open the RER daily on /rer/ → if it features Auber / Antony / CDG → their fame chips reflect the override values