feat(solver): surface bi-exponential fit outcome (FitMode)

[daharoni]

Problem

The bi-exponential fit could silently:

• fall back to slow-only (no fast component),
• produce a degenerate result — no positive slow amplitude, i.e. the free kernel had no real transient (noise/flat), or
• return the empty sentinel (0.02/0.4, residual = ∞) on empty input.

The caller had to infer which by inspecting beta_fast == 0 or an infinite residual. That's the second silent-fallback defensibility gap flagged in the config/numerical inventory (after the FFI NaN guard).

Change

Make the fit outcome explicit and carry it through the whole stack.

Rust (biexp_fit.rs) — add enum FitMode { TwoComponent, SlowOnly, Degenerate, Empty } and a fit_mode field on BiexpResult, classified once on the returned result:

• residual not finite → Empty
• beta <= 0 → Degenerate
• else TwoComponent / SlowOnly by fast-component presence.

FFI — WASM indeca_fit_biexponential serializes fit_mode (variant name) in its result object; PyO3 py_indeca_fit_biexponential appends it as a string (return is now an 8-tuple); Python BiexpFitResult gains a fit_mode field.

App (CaDecon) — the worker parses fitMode into KernelResult; iteration-manager counts Degenerate/Empty subset fits per iteration (degenerateSubsets / totalSubsetFits on the snapshot); KernelDisplay shows a vermillion ⚠ N/K fits degenerate warning when the displayed kernel rests on untrustworthy subset fits.

Tests

• Rust: Empty on empty input, Degenerate on a flat kernel, usable mode on a clean biexp.
• Python: test_fit_mode.py covers all three via the public fit_biexponential wrapper.
• Verified in-browser: a clean synthetic run shows the normal stats, no warning badge, 0 console errors.
• All green: 131 Rust, 32 Python, 289 TS; cargo fmt/clippy clean on both feature sets; typecheck clean; WASM builds.

Series

Config/numerical-hygiene series. Done: FFI NaN guard (#161), fit-mode status (this). Remaining (lower priority): expose result-changing tuning constants through config; de-dup + name the fast-grid bounds (latent drift hazard) + resolve the two biexp_fit.rs TODOs.

🤖 Generated with Claude Code