refactor(atlas): migrate to public warm_solve + golden the placement decisions

[jowch]

What

Now that MakieTextRepel shipped the public warm-start primitive we filed #24 for (PR #27, exported warm_solve), this does two things:

1. Migrate the Atlas off the internal-solver reach-through to the public API.
2. Golden the solver's placement decisions — newly possible because #G — CairoMakie Map Feature Page (examples/map_feature/) #27 guarantees+documents solver determinism.

Migration (placement-preserving)

• Atlas.jl: using MakieTextRepel: warm_solve (was ProjectionSolver, RepelParams, solve_cluster).
• place.jl: drop the _PARAMS/_SOLVER objects; call warm_solve(anchors, sizes, bounds; …, _SOLVE_KW...) with identical params. warm_solve is a thin face over the same seam.
• Project.toml: pin rev → 90f87ed3 (past #G — CairoMakie Map Feature Page (examples/map_feature/) #27/#I — README hero, gallery index, CHANGELOG, release hygiene #28). Stale solve_cluster mentions in comments updated to warm_solve.

Verified placement-preserving: a byte-identical hero re-render (same sha256, 241,054 bytes) plus the unchanged geometric golden. So the public API produces exactly the same placement — zero behavioral change.

Placement golden (machine-stable)

The existing geometric golden deliberately avoids solver offsets because they flow through Makie's projection, which isn't bit-stable across versions. So instead of hashing raw offsets, placement_rows/atlas_placement_digest hash the solver's discrete decisions — which side each label leans (offset quadrant) + dropped — from the real cold solve:

• Faithful to the actual map, but a quadrant only flips on a genuine re-placement, so it stays machine-independent like the geometric table.
• Offset magnitude / frame-to-frame continuity remains covered by test_loop.jl's warm-start delta bound (<25px/frame).
• New baseline test/golden/atlas-placement.sha256 (+ .rows.txt), UPDATE_GOLDEN=1 to regenerate.

Tests

Full suite 295/295. (Manifest stays gitignored; CI's clean instantiate fetches the new rev.)

Notes

• worktree demos C2 shape pack multi #25 (recipe measure/solve split) and [WIP] #E — Asteroid TUI demo (interactive polish in progress) #26 (measure_labels export) don't affect the Atlas — it uses warm_solve directly and measures via TextMeasure. With #H — CairoMakie "Newer Yorker" correctness exhibit (examples/cover/) #24 resolved, the Atlas now depends only on public MakieTextRepel API.

🤖 Generated with Claude Code

[jowch]

Review — round 1

Clean, behavior-preserving refactor plus a well-constructed placement golden. The solve_cluster(_SOLVER, …) → warm_solve(…) migration is exact: warm_solve is a thin pass-through to the same solve_cluster(ProjectionSolver(params), anchors, sizes, bounds; init_state, pin_mask, pinned_offsets, obstacles) seam, and place.jl passes its solver params (box_padding=5.0, point_padding=5.5, min_segment_length=2.0, only_move=:both) explicitly, so it's insulated from any drift in warm_solve's own defaults. The new golden hashes discrete side/dropped decisions (sign-quadrant + Int(dropped)), not pixels — deterministic and meaningful. I independently reproduced the committed 475dfc52… sha256 from atlas-placement.rows.txt and ran the full atlas suite (295/295 pass) against the committed sha. Measure-once/layout-many and the no-kerning rule are untouched. No blocking findings.

Nits

• examples/atlas/src/golden.jl:115 — placement_rows' determinism silently depends on the mutable global _PIN_SLO[] (render.jl:106) staying true. Nothing in the test path flips it, but passing the pin flag explicitly (or a one-line comment noting the dependency) would harden the golden against a future test that mutates the Ref.
• examples/atlas/src/golden.jl:14-17 / examples/atlas/test/test_golden.jl:43-45 — the "machine-stable" framing leans on the same "needs no Makie" argument as the geometric atlas_digest, but placement_rows routes anchors through the live Makie camera (assemble_frame → _data_to_px → Makie.project). The digest is robust to sub-pixel projection drift, not independent of Makie — a label near a quadrant boundary could flip under a projection change. Wording should say "robust to sub-pixel projection drift" rather than implying full Makie-independence.
• examples/atlas/README.md:28,55,68 — still reference solve_cluster; now contradicts place.jl's public-warm_solve reality. Minor doc drift (the archival prep/spike_timing.jl keeping the old name is fine).

[jowch]

Round 1 — addressing the nits

Pushed 0f33077. All three valid; fixed all three. Atlas suite 295/295, placement digest unchanged.

Fixed

• _PIN_SLO[] dependency (golden.jl) — placement_rows now explicitly pins _PIN_SLO[] = true (the production value assemble_frame reads) before solving, so a future test that leaves the global Ref flipped can't perturb the golden. It was already true, so this is a no-op on the digest (baseline 475dfc52… unchanged) — pure hardening.
• "machine-stable" overclaim (golden.jl header + docstrings, test_golden.jl comment + testset name) — you're right: placement_rows routes anchors through Makie.project via assemble_frame, unlike the fully-affine geometric table. Corrected throughout to "robust to sub-pixel projection drift" (a quadrant only flips on a real re-placement), and made explicit that it is not Makie-independent like the geometric one. The robustness argument stands; the wording no longer overclaims the mechanism.
• README solve_cluster drift — updated lines 28/55/68 to warm_solve, matching place.jl. (Left prep/spike_timing.jl as the archival record of the original spike, per your note.)

[jowch]

Review — round 2

All three round-1 nits addressed in 0f33077, no regressions. Scoped post-approval re-review of the incremental diff (golden.jl, test_golden.jl, README.md).

The only functional change is _PIN_SLO[] = true at the top of placement_rows (golden.jl:113). I verified it's inert: _PIN_SLO defaults to Ref(true) (render.jl:106), golden.jl:113 is its only writer (and writes true), and it's read only at render.jl:581 — so there's no ordering hazard and the digest is unchanged. The golden data files (atlas-placement.rows.txt/.sha256) are untouched in this diff, so the 475dfc52… digest already verified in round 1 still holds by construction. The remaining changes are comment/docstring/testset-name wording, now correctly stating "robust to sub-pixel projection drift, not Makie-independent."

Resolved since last round

• _PIN_SLO[] dependency — now pinned explicitly in placement_rows; confirmed inert (sole writer, default already true). Fixed.
• "machine-stable" overclaim — corrected throughout golden.jl header/docstrings and test_golden.jl to "robust to sub-pixel projection drift," with the Makie-routing caveat made explicit. Fixed.
• README solve_cluster drift — lines 28/55/68 updated to warm_solve; archival prep/spike_timing.jl left as-is per the note. Fixed.

[jowch]

Converged at round 2 — zero blocking throughout; the three round-1 nits (the _PIN_SLO hardening, the "robust to sub-pixel projection drift" wording correction, and the README solve_cluster→warm_solve drift) are fixed and re-approved. Migration is verified placement-preserving (byte-identical hero, unchanged geometric golden); the new placement golden holds at 475dfc52…; suite 295/295. No further changes intended.