fix(descent-bench): route the region-conditioned arm through the real edit seam (+5 hardening riders)

[eren23]

The headline: the descent bench was measuring a fiction

The descent bench's region-conditioned "with" arm called generate_image with the region crop as a reference_url. But fal's text-to-image endpoints silently ignore reference_urls (accept-but-drop, same as the fresh-gen nano endpoints per PR #109). So the arm that was supposed to prove region-conditioning helps was actually unconditioned — a fresh generation wearing a reference URL — and place_lift sat structurally ~0.

Fix: route the with-arm through the seam the product actually ships, with the crop as the edit SOURCE so reference pixels bite:

• interior chains → build_enter_instruction + edit_image on the enter_scene model
• exterior closeups → build_zoom_instruction + continue_image (Kontext), the closeup rung

Two regression guards (test_with_arm_rides_edit_seam_not_generate, test_exterior_with_arm_rides_zoom_continue) spy on which seam fires, so the with-arm can never silently slip back to generate_image.

Five hardening themes riding along

1. Coordinate-scale ladder → shared module. Extract the fraction/percent/per-mille coercion out of llm/click.py into providers/coordinate_scale.py and wire it into detector.py + llm/extraction.py (was click-only). Closes the long-standing "world extraction/detector have the same scale flip" coverage loss — pixel-looking coords now drop instead of edge-clamping.
2. Enter-retry model swap (off by default). ENTER_RETRY_MODEL_SWAP keeps attempt 0 on the router's production pick and swaps only judged retry attempts to a cheaper nano-banana-2/edit, threaded via a new render_for_attempt hook through render_loop → tap + both benches. An explicit request-level image_model override is respected on every attempt.
3. Judge transport hardening. Pillow-downscale the VLM judge payload (VLM_JUDGE_MAX_SIDE_PX/VLM_JUDGE_JPEG_QUALITY) and route the judge call through _create_with_retry. Multi-MB bench JPEGs were large enough inline to reset OpenRouter/Gemini reads; falls back to the original bytes if Pillow can't decode.
4. Web: disable prefetch/precompute warming in World Mode. World taps route through geometry / cold resolver context, but the precompute effect was seeding candidate-only warm entries carrying stale enter_as/place_form classifications. worldEnabled now early-returns from both the precompute effect and the hover-prefetch handler.
5. Bench ergonomics. Artifact reuse (*_REUSE_ARTIFACTS), arm/case filters (VIEW_BENCH_ARMS/VIEW_BENCH_CASES), style-ref compaction before fal upload, judge retry+backoff, positioning-probe toggle.

Env docs updated in both .env.example files; conftest.py scrubs the two new env vars.

Verification

• Backend: 983 passed, 2 skipped, coverage 86.71% (floor 85; new coordinate_scale.py + model_router.py at 100%).
• Web: tsc --noEmit clean; pre-commit eslint --fix + ruff --fix clean.

Not in this PR

The live descent-bench run that empirically proves place_lift now lifts post-fix is not included here — it costs real fal image + VLM judge spend. Happy to run it as a follow-up (with DESCENT_BENCH_REUSE_ARTIFACTS off for a true measurement) once someone confirms the fal balance.

🤖 Generated with Claude Code

[cursor]

Bugbot is not enabled for your account, so this pull request was not reviewed.

Enable Bugbot in the Cursor dashboard to get automatic reviews on future PRs.

[eren23]

Live descent-bench run (n=1, ~$0.30)

Ran make eval-descent (DESCENT_BENCH_RUN=1). One golden chain resolves: village-medieval-manor :: 'Church' -> wm-chester-cathedral-nave.

Result: mean place_lift: +0.000 (style_lift: −7.500), continuity_with: 9.0, both arms place_match = 2.0.

The seam fix is verified working. Logs confirm the with-arm rode image.edit (fal-ai/nano-banana-pro/edit, refs conditioning) and the without-arm rode image.generate (refs: 0) — no more generate_image-with-ignored-refs. continuity_with = 9.0 proves the region conditioning actually bit: the generated interior is coherent with the cropped map region.

Why place_lift is still 0 here — the chain is degenerate, not the fix. Verified by eyeballing all four artifacts:

• region input = a vintage village map where "Church" is a tiny cross glyph (a small country parish).
• with-arm = a faithful illustrated small-church interior (stays in the map's medium — correct descent behavior).
• without-arm = a photoreal small parish church.
• real child = a photo of Chester Cathedral — a colossal Gothic cathedral.

The map never depicted Chester Cathedral, so neither arm is the same place → place_match = 2.0 for both is the honest, correct score. style_lift = −7.5 is by-design (the with-arm preserves the illustrated medium, which is exactly why place_match is the medium-agnostic headline).

Takeaway: the branch's real achievement — the bench now measures the true product seam instead of an unconditioned fiction — is confirmed. Demonstrating place_lift > 0 needs a golden chain whose real child photo actually corresponds to a distinctive place depicted in the parent map (corpus work, out of scope for this PR).

Artifacts (gitignored) in tests/map_corpus/overlays/descent-wm-chester-cathedral-nave-{region,with,without}.jpg.