Incremental steady-state fit driver (capstone of #419)

[matthewholman]

Summary

The capstone of incremental orbit determination (issue #419): incremental_orbitfit, a single steady-state maintenance pass over a batch of objects that ties the three levers together. For each object it routes:

• skip — the observation set is unchanged since prior_catalog (matching fingerprint): carry the prior fit forward verbatim, no fit.
• sequential update — observations were only appended (prior_obs given and every prior observation still present): update via sequential_update over the new observations only. Its nonlinearity gate falls back to a full refit when the update is too large.
• full refit — observations removed or changed, or no per-object prior_obs: refit over all current obs, warm-started from the prior state.
• cold — no prior for the object: cold IOD fit.

Returns the updated result catalog (same schema as orbitfit output, carrying the obs_hash/nobs_fit fingerprint so it seeds the next cycle) plus a routing tally.

Obs-level diffing (_obs_row_keys / _append_only_new_obs) reuses #420's per-row fingerprint strings, so a row's diff key equals its contribution to the object's fingerprint — append-only vs modified is detected exactly.

⚠️ Stacking — review order

This is one feature in three parts and stacks on #420 and #421:

• Incremental fits: obs fingerprint + skip-unchanged (Phase A of #419) #420 — obs fingerprint + orbitfit(skip_unchanged=True) (the skip route + schema).
• Sequential / information-filter orbit update (Lever 3 of #419) #421 — the sequential / information-filter update (run_sequential_update + sequential_update).

This branch merges both, so the diff below includes them. Merge #420 and #421 first; this PR's net-new code is a single commit, 091f804 (incremental_orbitfit + tests/layup/test_incremental_driver.py) — review that commit directly, and I'll rebase once the two land so the diff reduces to just the driver.

Validation

• 5 route tests in test_incremental_driver.py (skip byte-identity, sequential-matches-full-refit, modified-obs→full, cold, no-prior-obs→full). Full suite of Incremental fits: obs fingerprint + skip-unchanged (Phase A of #419) #420 + Sequential / information-filter orbit update (Lever 3 of #419) #421 + this + base fit tests: 31 passed.
• Real MPC data (mpc_demo/cycle_rehearsal_seq.py, a steady-state cycle on numbered asteroids): the sequential route matches a full refit to state rel ~1e-8 / KEP a ~1e-9, at 6.5–13.7× less wall-clock on ~1000-obs arcs (only the new obs are integrated; the win approaches the integration-work ratio as the arc lengthens). The nonlinearity gate never tripped on well-observed objects with small nightly additions — its job is the pathological cases.

Combined with the skip route, the steady-state loop is: 95–99% of objects skip (zero cost), changed objects sequential-update at ~7–13× less work, and only non-append or nonlinear changes fall back to a full refit — the operational maintenance loop JPL/MPC run.

Follow-ups (unchanged from #421)

Outlier rejection on the new observations in the sequential path; non-grav priors; radar/occultation in the sequential path; differing-epoch priors. A CLI wrapper for incremental_orbitfit can follow once the API settles.

Part of #419.

Closes #419

[kjnapier]

Looks good to me