docs(handoff): restore HANDOFF_HEAVY_TEST_GAPS.md with investigation update

The handoff was removed in e9837a0 ("Remove HANDOFF_HEAVY_TEST_GAPS.md
per user request") earlier this session. Restoring the original content
from 3facdec and appending a session-2 investigation update with
concrete root-cause data for each gap:

Gap 1 — FIXED in 03bf3f8 (CLLLC netlist precision: std::fixed was
rounding sub-microscale Cr / maxStep values to "0.000000" and
ngspice rejected the netlist syntactically). [clllc-topology][
ngspice-simulation] tests go from 2 SKIPPED → 2 cases / 14 assertions
all pass. WebFrontend `runSimulated` step should complete instead of
the previous 180-360 s timeout.

Gap 2 — `[!silent-shortcut]` 60 s wall-clock budget attempt
(commit 4b4ebc4) was reverted in 0fd66d0 per the CLAUDE.md "no
fallbacks, no silent shortcuts — throw" policy. The right fix is
the structural one the handoff originally suggested: candidate-
budgeted top-K per winding in the inner wind loop, capping at K^N
instead of |wireDatabase|^N. Needs a deterministic MKF-side
reproducer first (capture the actual MAS JSON the WebFrontend
posts to `calculate_advised_coil` and write a unit test against it).

Gap 3 — Concrete diagnoses from running PSFB through MagneticAdviser
with Logger::setLevel(INFO):

  PSFB ferrite path drops 168 → 0 at the **Saturation** filter.
  All 168 ferrite cores that satisfy the inductance constraint
  exceed the 1.2× Bsat margin at PSFB's V·s peak with the
  initial-N derived from `add_initial_turns_by_inductance`.

  PSFB powder path drops 4000 → 0 at the **Inductance** filter.
  Powder cores' lower max permeability can't reach PSFB's
  requested Lm at any reasonable N for the dimension-passing
  shapes.

  Underlying issue: PhaseShiftedFullBridge.cpp:220-226's
  Im_target=0 path forces `Lm = max(Lm, 20 * Lr)` which for
  600 W at 12 V × 50 A drives Lm high enough that initial-N
  selection puts B_peak past Bsat for the entire catalog.

  PFC at 50 W takes 5:25 in standalone MKF; CoreAdviser is the
  bottleneck (CoilAdviser is fast for single-winding PFC).
  Needs per-stage time profiling to target the specific
  filter — most likely `add_ferrite_materials_by_losses`
  sweeping the candidate set × operating points.

Gap 4 — likely cascades from Gap 2 + Gap 3 fixes.

Recommended fix order for next session is captured in the doc;
none of the gaps are fixed in this commit, but the diagnostic
work is now actionable instead of speculative.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Author: AlfVII