0.9.0 — alpha flatten, placement envelope, score presets

Changed

• Transparent inputs flatten before decode (pipeline 1→2,
  score_contract 3→4) — the engine used to read the RGB stored under
  the transparency (alpha was dropped at the input seam), which made the
  verdict exporter-dependent: the same visual design scored 100
  (white-under-alpha exports), 74 (canvas-style black-under-alpha) or
  "no detection" (dark modules over canvas transparency) depending on
  which tool wrote the PNG. Since this release the input composites over
  a DECLARED background before luma conversion and RGB extraction —
  every stage (ladder · score · UEC · ISO) sees the flattened image, and
  the stored under-alpha RGB can never influence the report
  (exporter-invariance is pinned by test). auto (the default) measures
  the design on its intended placement: dark content over white, light
  content over black, with an opposite-background retry on a
  zero-detection scan (alpha.fallback_used) so a mis-called mean can
  never produce a false "no detection". Opaque inputs are untouched —
  their reports stay byte-identical (the alpha key is omitted
  entirely, not null). Spec: 01-report § Alpha · 07-pipeline.

Added

• ScanConfig.alpha_background on every surface, per the
  score_skip_axes precedent — core (AlphaBackground +
  from_name), wasm (scan_image(..., alpha_background)), Node
  (alphaBackground option, typed
  "auto" | "white" | "black" | "none" | hex), CLI
  (--alpha-background), Python (alpha_background= kwarg, stub
  updated), UniFFI/Kotlin/Swift (alpha_background arg with a typed
  UnknownAlphaBackground error). #rrggbb forces the host's real
  placement surface — the truest verdict when the embedder knows where
  the code will sit; none restores the drop-the-channel behavior as
  an escape hatch. Unknown values fail loud on every surface. (The
  Flutter facade parses the new block; its option arg lands with the
  facade's next options pass — it exposes neither budget nor skips
  today.)

• The placement envelope (alpha.envelope, Full profile) — the
  transparent-design question is not "does it scan" but "over which
  backgrounds does it keep scanning": five fixed neutral rungs
  (luma 0 · 64 · 128 · 192 · 255) plus one bisection probe per verdict
  boundary, each a quick decode in the primary symbol's own decode
  class. The report carries the probes (every endpoint a TESTED
  background, never interpolation), the contiguous safe_luma bands,
  and a placement verdict (any · light_only · dark_only ·
  mixed · none). Informational by contract: it never moves
  score.value. Skippable through the standard seam
  (score_skip_checks: ["alpha_envelope"] — ScoreCheck gains the
  variant on every surface).

• Hint::AlphaBackgroundDependent { placement } — fires when the
  envelope is narrower than any: pin a background layer in the
  artwork, or constrain placement to the safe bands. The generator/UI
  loop's alpha feedback.

• report.alpha block: coverage · content_luma · mode ·
  background ("white" / "black" / "#rrggbb") · fallback_used
  · envelope. Schema, TS types, Dart mirror and the type-parity gate
  updated in lockstep; qrscan --pretty renders the flatten verdict
  and the envelope strip.

• alpha_palette on every option surface (alphaPalette /
  --alpha-palette) — the host's theme/brand colors probed by the
  envelope inside the SAME scan: per-color verdicts land in
  alpha.envelope.palette (request order, each at its BT.601 luma),
  instead of one full scan per candidate background. Palette colors
  never distort the neutral probes/bands/placement. Unknown colors fail
  loud (white · black · #rrggbb only — the flatten keywords are
  modes, not colors).

• Hint::AddBackgroundPlate { color } — the canonical REMEDY riding
  with the alpha_background_dependent diagnosis: add an opaque plate
  (rounded rectangle + quiet-zone margin) behind the symbol; color
  (white for dark content, black for light) maximizes contrast.
  Maps 1:1 to an editor's one-click fix.

• score.weights_run — the honesty integer: Σ contract weights of the
  axes that ran (100 = the full six-axis contract). A partial score now
  SAYS how much contract stands behind it; pre-0.9 reports parse
  leniently (default 0). qrscan --pretty renders "· 70% of contract"
  when partial.

• score_preset on every option surface (scorePreset /
  --score-preset): the two canonical postures by NAME — design
  (generated previews: skips perspective+rotation, keeps lighting) ·
  capture (all six). Sugar over score_skip_axes, mutually exclusive
  with an explicit list (rejects loudly on every surface) — N host
  integrations can no longer drift apart on hand-built lists.

• axes[].refined_failed_at — the knee, bisected once at Full depth:
  the tightest TESTED failing intensity ("blur 2.25", "435px", "63%").
  One extra probe per kneed ramp; the composite never reads it; absent
  when no refinement ran. The generate→scan→regenerate loop sees
  sub-cell progress instead of 2-4 point cliffs.