feat(records): ZeroDimConfig.recall — force a fresh re-track past the records recall

[ofloveandhate]

ZeroDimConfig.recall — force a fresh re-track past the records recall

A recording solve recalls an identical ask (same system + settings + seed) instead of re-tracking
it (ADR-0046) — correct, and usually what you want. But it silently starves anything that needs the
tracking to actually happen:

• a path-collecting observer (SolutionPathCollector) comes back empty — you attached it, solved,
  and got nothing, with no error;
• a benchmark measures recall time, not compute time;
• a re-verification of a recorded run never actually recomputes.

This came up live: a path-portrait observer collected 0 paths on a re-solve, and it took a full
diagnostic pass to realize the paths were recalled, not tracked. Recording working — but confusing.

The fix

Add ZeroDimConfig.recall (default True). Set it False to force a fresh track of every path
even when the paths are recorded; the fresh run is still recorded.

cfg = solver.get_config(ZeroDimConfig)
cfg.recall = False          # re-track even if this exact ask is already recorded
solver.set_config(cfg)

Like num_threads, it's transient — it changes only whether/how the work runs, not what is
computed — so it is deliberately excluded from the config's canonical encoding / identity digest (the
digest is unchanged; no version bump).

(There is a second, seed-based way to dodge recall: seed = 0/None draws fresh entropy each solve, so
the ask is never identical twice. That's the "give me a fresh random run" lever (non-reproducible);
recall=False is the "recompute THIS exact reproducible ask" lever. Different intents — this PR adds
the latter.)

Changes

• gate RecallRecordedPaths on the flag in the shared solve loop (zero_dim_solve.hpp)
• exclude recall from CanonicalEncoding(ZeroDimConfig) (mirrors num_threads)
• expose config.recall in the binding
• C++ test: recall=false re-tracks an otherwise-recalled ask (NumPathsRecalled()==0, roots still found)
• Python test: recall=False re-tracks so a SolutionPathCollector collects again

Verification

• test_nag_algorithms + test_classes green (config-digest fixture unchanged).
• pytest python/test/zero_dim/ — 121 passed.
• doclint not run locally (no doxygen in env); the new recall field carries a Doxygen comment.

🤖 Generated with Claude Code