Record active phi-grid placement decisions

[krystophny]

Summary

This PR adds an opt-in record of the phi-grid strategy that modify_propagator actually applies.

The first implementation recorded only automatic phi_placer intervals. A production reconstruction then emitted no CSV because its configured path uses phi_split_mode=1; the automatic eta placer is never called. The recorder now covers the decision boundary shared by all modes:

• placed: automatic eta-interval placement, with the raw CEILING decision and the count after the existing odd/minimum adjustment;
• halfstep: configured mode-1 refinement, with input and output grid-point counts;
• no_eta_halfstep: the mode-2 fallback when no eta value lies in the propagator interval;
• divider: the mode-3 retry path, with input and output grid-point counts.

Each row identifies the propagator and outcome. It also records the interval, branch, eta index, phi endpoints, and original phi spacing. The generic count columns are named decision_count and result_count because their exact meaning is strategy-specific.

Behavior and invariants

Set NEO2_PHI_PLACEMENT_FILE to request the CSV. With the variable unset, the recorder performs no allocation or file I/O. A requested path that cannot be written is a hard error rather than an incomplete record.

The diagnostic does not change the selected strategy, CEILING expression, parity adjustment, half-step or divider grid, point locations, field-line integration, source, collision operator, join normalization, convergence criteria, ABI, or generated physics-data layout. It only observes the grid before modify_propagator replaces the old coordinates.

Relation to the flux-pumping stack

This branch is stacked on PR #154 because the downstream convergence campaign uses one pinned executable history. It does not depend on the join-normalization values. The record is intended to distinguish a response change caused by a phi-grid strategy or count transition from a downstream change.

Verification

Failing before the recorder existed:

phi_placement_diagnostic_test: record_phi_placement is absent

The first production attempt at the previous head also failed strict validation because no requested CSV was created. That exposed the process-coverage gap: the unit test exercised the writer, but the carrier never entered automatic placement.

Focused test at defe172:

phi_placement_diagnostic_test  PASS  0.06s
Summary: 1 passed, 0 failed, 0 skipped

Full repository pipeline:

Static: OK (348 modules, 348 changed, 348 affected)
Build: OK
Tests: OK
Lint: OK
All stages passed
Fmt: WARN

The format warning is existing whole-file drift in COMMON/flint_mod.f90 and NEO-2-QL/flint.f90; the touched lines and staged diff pass whitespace and line-length checks.

A real three-stage reconstruction at nstep=480 now writes 203 validated rows in stage 0 and 203 in stage 2. The records are byte-identical between those stages: 199 use halfstep and four use the divider retry path. Decision counts span 5-865 grid points and result counts span 9-1739. The solve retains 199 reconstructed responses and closes qflux below 8.7e-15 relative error.