A tiled image born on a placeholder rendered solid black

[CSSFrancis]

Closes #60.

Why

A tiled imshow that entered tile mode on a flat placeholder (zeros, before real
data exists) derived the fixed quantisation band raw_min/raw_max from that
placeholder — a degenerate (0, 0) — and nothing re-derived it on a later
set_data. The two ends of the protocol then disagreed about what (0, 0) meant:

• the Python encoder (_tile_quant_clim) treats a degenerate band as unset and
  quantises the tile bytes over the display window instead;
• the JS LUT honoured it literally — hMin + raw/255 * (range||1) over a zero range
  puts all 256 codes in [0, 1], below any sane display floor.

Every code clamped to colormap[0]: a 256-entry black LUT, on the WebGPU shader path
and the Canvas2D blit alike. Stats and histograms bypass the LUT, so they looked
perfectly healthy next to a black pane, and nothing warned. Downstream, every 8k still
opened into a figure laid out on its zeros placeholder showed this.

What

Both ends, as the issue suggested:

• Python — extract the guard enable_tile already had into _ensure_tile_band()
  and call it from update_tile_source, the single seam every data swap goes through,
  so set_data and the public update_tile_source are both covered. A band that is
  already valid is still never re-derived, so a contrast change keeps re-windowing in
  the LUT with no pixel re-encode.
• JS — read the band through one _rawBand(), shared by _buildLut32 (both
  render paths) and by the colorbar tick placement, which read the band the same way
  and was equally wrong.

One correction to the issue's suggested JS rule: mirroring _tile_quant_clim
unconditionally is wrong, because a degenerate band means two different things. On
the plain path _normalize_image emits all-zero bytes and reports (c, c) — the band
names the constant value code 0 stands for, and the old reconstruction was correct.
The unconditional version turns every constant untiled image black; it broke seven
layer-blend tests. The fallback is gated on tile_enabled.

Also adds set_tile_band(vmin, vmax) — the issue's third ask. The fix alone does
not let a host drop its live-camera workaround, since that pins a known range rather
than a data-derived one, and a valid band is deliberately never re-derived.

Tests

12 new cases in test_tiled_imshow.py, including two Playwright ones. Verified to
fail without the fix — the browser test reproduces the exact symptom,
{'lo': 0, 'hi': 0, 'mean': 0}.

One existing test changed. test_set_data_stays_identical_between_plain_and_tile
asserted byte-identical plain/tiled payloads, but only passed because of this bug:
on unmodified main, giving it a real base instead of zeros fails at 161 differing
bytes, since a tiled plot quantises over its band and a plain one over the clim. It is
now parametrised over both bases and asserts what the viewer actually sees — identical
within one colormap step (measured: max |diff| = 1). The wire-byte guarantee at
construction is untouched and still passes.

2144 passed, 58 skipped.

Notes for review

• The behaviour change is filed as api_change per AGENTS.md: a tiled plot born on a
  placeholder now quantises frames over the frame's own range rather than the display
  window, matching what imshow of a large frame has always done. Wire bytes can
  differ by a rounding step from the equivalent untiled plot; the displayed image does
  not.
• FIGURE_ESM.md anchors were already 42 lines stale from an earlier commit, so all 55
  were re-measured against the source rather than blind-shifted, and the verified line
  count in AGENTS.md was corrected.

[codecov-commenter]

Codecov Report

❌ Patch coverage is 95.45455% with 1 line in your changes missing coverage. Please review.
✅ Project coverage is 90.90%. Comparing base (e538884) to head (c6626df).

Files with missing lines	Patch %	Lines
anyplotlib/plot2d/_plot2d.py	95.45%	1 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##             main      #62      +/-   ##
==========================================
+ Coverage   90.88%   90.90%   +0.01%     
==========================================
  Files          40       40              
  Lines        4532     4550      +18     
==========================================
+ Hits         4119     4136      +17     
- Misses        413      414       +1     

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.
• 📦 JS Bundle Analysis: Save yourself from yourself by tracking and limiting bundle sizes in JS merges.