Budget placement pressure and localize the linearize witness

[fpedd]

placement_pressure_per_allocation ran its internal linearize attempt
unbounded, so large non-interval-order vector-clock instances paid a
futile O(kmd) dominance pass before the fast output-sensitive
fallback sweep. Plumb work_budget through: Python defaults to
DEFAULT_WORK_BUDGET, and the budget bounds the linearize attempt
(silent fallback, as in omni_place) as well as the fallback conflict
sweep (RuntimeError past it, matching conflicts/conflict_degrees).

The 256-sample random incomparability witness in the linearizer missed
temporally local concurrency (global samples, local 2+2s), so genuinely
concurrent instances fell through to the full O(kmd) chain test. Add
a deterministic O(k*(log m + d)) local pass probing every
lexicographically adjacent start-row pair against the end-row windows
just below the component-0 boundary; every hit is a sound 2+2 and a
miss still reaches the exact chain test, so results are unchanged. The
budget check now runs before the witness so refused instances skip the
scan.

Measured: omni_place(linearize_budget=None) on 1M loosely coupled
vector allocations drops 162 s -> 1.3 s, and placement pressure on a
400k near-chain instance with one hidden 2+2 now answers in 0.2 s.
Differentially verified against brute-force references (conflict
relation, pinned antichains, placement peaks, interval-order decisions)
on over a thousand randomized instances with zero mismatches.