fix(core): correct Crucible-confirmed QS + spectral physics defects

[jungdaesuh]

Summary

Fixes five confirmed correctness/contract defects in the QS + spectral core, surfaced
by an adversarial (Crucible) review. Each is covered by a regression test that fails
against the pre-fix behavior. No behavioral scope creep.

Related issue

Relates to #40 — three of these are in the pure core/quasistationary port, a
prerequisite to wiring it in place of the _slcontour reference.

Type of change

• Analysis core (src/magnetics/core)
• Data source (src/magnetics/data)
• Service / API (src/magnetics/service)
• Web GUI (gui/web)
• Docs
• Other

What changed

• qs_bridge: fit-quality status vocabulary (ok/warn/error) diverged from the
  contract's good/warn/bad, so the GUI traffic light lost its color for healthy /
  ill-conditioned fits → route through contracts.quality_for_k (SSOT).
• _slcontour/fit.py + core/quasistationary.py: per-coefficient SVD σ summed the
  wrong axis (axis=0 → 1/w_k² by singular position, not the covariance diagonal) →
  wrong live QS error bars; axis=1 also sizes fit_sigmas to n_cols, fixing an
  underdetermined-fit IndexError.
• _slcontour/fit.py: the 2D sinusoidal-integral basis divided by
  deg2rad(dx*dy)*n*m (π/180 once, no i²) → m≠0 columns off by −180/π; use the
  separable product of the two 1D branches.
• core/quasistationary.py reconstruct_grid: used c instead of conj(c) vs its
  own forward model → toroidal mirror.
• core/spectral.py: phase_sigma reported the inverse-variance (BLUE) variance
  while the intercept is amplitude/coherence-weighted → propagate σ through the same
  weights (honest, non-optimistic 1σ).

Note: the review's initially-reported except-syntax "critical" was a false
positive (valid PEP 758 on the pinned Python 3.14). No change here.

How I tested it

uv sync --extra service; uv run ruff format --check . / ruff check . /
ty check src/magnetics → clean; uv run --with pytest pytest -q → green
(tests/test_crucible_regressions.py + a test_qs_bridge quality-vocab test, each
failing against the pre-fix behavior).

Checklist

• CI is green (analysis: ruff + pytest; web: lint + build) — validated locally
• No secrets/credentials committed
• No large files or tokamak data committed
• Physics kept in core (device-agnostic); none added to service routes
• Docs updated if behavior/usage changed (n/a)

🤖 Generated with Claude Code

[matt-pharr]

Superseded by #68, which is a rebased superset of this PR on current develop (contains all five fixes here: the SVD-σ axis=1 fix, the 2-D sinusoidal-integral denominator, the conj(c) reconstruction, the quality_for_k status vocab, and the weighted-σ phase_sigma). Closing as a duplicate — review/merge #68 instead.