feat: zone placement — pin tones to zones, solve density and grade

[marcinz606]

What

The spot densitometer made actionable — what a darkroom enlarging analyser (RH Designs Analyser Pro) does. A new crosshair tool on the canvas toolbar:

• Click pins a spot (max 2; a third click replaces the nearer pin). Each pin shows on the canvas as a numbered ring with its measured zone.
• The Analysis panel grows a row per pin: what it reads now, and a target zone stepped in thirds (Roman numerals).
• One pin → Place zones solves Print Density so that tone prints on its target. Two pins (shadow + highlight) → solves Print Density + ISO-R Grade so both land.
• Stepping a target live-previews the solve without committing (config-override render). Apply commits one undoable edit and turns off Auto Density (and Auto Grade for two pins) — a meter left on would re-move the placed tones.
• Asking for a zone the paper can't reach shows an amber → lands … with the closest achievable zone; the solve pegs at the slider bound like an analyser pegging at grade 5.
• Pins are proofs, not edits: Esc clears them (second Esc puts the tool down), any other edit or frame switch drops them.

How

• Pure solver negpy/features/exposure/placement.py: bisection over the chart's own print_curve (green/base channel) via curve_params_from_metrics → print_curve_output → zone_of_encoded, so measured/target/achieved sit on one ruler with the chart and step wedge. Two-pin is a nested bisection (outer grade, inner density pinning the dark tone). Bisection depth matches slider resolution (0.01 density / 1 R) — 2-pin solve ~33 ms.
• encoded_of_zone added beside zone_of_encoded (exact closed-form inverse, same file, so the ruler can't fork).
• Pins ride the test-strip lifecycle: dropped by any real render / load_file, kept through the one render Apply fires. Patch sampling is a 5×5 mean of normalized_log through a _sample_normalized_log helper factored out of the hover probe (hover keeps its 1×1 read).
• Split grade folds into every candidate curve, so the two-pin solve stays enabled with trims active. B&W/E-6 need no branch.
• No render-path change: solver writes existing ExposureConfig fields; no CPU/GPU/WGSL work, no parity surface.
• Shortcut registry entry ships unbound (every Tools-row Shift+letter is taken); bind in the shortcut editor.

Docs

USER_GUIDE gains a Zone placement subsection in the Analysis panel section (feeds the in-app ⓘ guide). PIPELINE.md untouched — no pixel math changed.

Verification

• 25 new tests (tests/test_zone_placement.py): solver landing/clamping/degenerate/split-grade/rounding, controller pin lifecycle, sidebar widget. TDD, each watched fail first.
• make all green; full suite 2947 passed.
• Headless E2E on the real app (samples/DSC00448.ARW): one-pin landed 8.326 vs target 8.333; two-pin landed 9.00/3.35 vs targets 9.00/3.33 with grade solving to 159 R; autos, single undo step, Esc grammar and pin ephemerality all asserted.

Update — pins are draggable, and read live

• Press a pin to grab it (16 px hit radius, hand cursor on hover, crosshair for the tool itself); only a press on bare frame drops a new pin.
• The tone under a dragged pin is re-read as it moves, so its zone tracks the cursor. Caption reads 1 · IV⅓ → VI — where the pin reads now, and where it is asked to print; the arrow drops once the two agree.
• A pin still on its own reading follows the tone it lands on; one whose target has been stepped keeps that target through a drag.
• The solve waits for the drag to end (two-pin nested bisection is ~15 ms, too slow per mouse-move); the last solution stands in for the sidebar meanwhile.
• 13 more tests (tests/test_zone_pin_drag.py + controller lifecycle cases); headless E2E drives real Qt press/move/release on the overlay: pin placed, dragged, zone re-read (IX⅓ → VI⅔), target retention verified, applied tone landed 8.326 vs target 8.333.

Update 2 — the zone strip is the control

Entry point redesigned: the toolbar crosshair button and its (unbound) shortcut are gone.

• Click a zone on the Analysis zone strip, then click the photo — that spot is asked to print on that zone. Clicking the armed cell again cancels; the armed cell is outlined and a HUD hint says what the next click will do.
• Placing previews the solve at once. Place zones (now the panel's primary/accent action) or Enter over the canvas commits and closes the tool; Esc discards the armed zone, then the pins and the preview with them.
• One pin is enough to accept. Each row has its own ✕; removing the last pin rolls back to the committed print.
• A canvas click with nothing armed still just meters (pin at its measured zone, print untouched).
• Zone strip height matched to the step wedge; ToolMode.ZONE_PLACE stays internally (crosshair cursor, pin hit-testing, drop-pins-on-exit).
• Tests: strip click→zone mapping, arming/disarming lifecycle, armed vs unarmed clicks, per-pin removal, Enter-accepts (1 and 2 pins), Esc ladder. Full suite 2974 passed. Headless E2E drives real presses on the sidebar strip and canvas: arm → place → remove → re-place → Enter → committed solve matches the sidebar's promise (clamped ask reported as amber "lands"), then the Esc ladder.