[fpedd]

Every allocator now reads the conflict relation as one CSR adjacency instead of unpacking it into a vector per allocation, and one place_order kernel serves first fit, best fit, the omni portfolio and FirstFitPlacer, replacing four copies. The greedy orders, including greedy_by_contention from #78, are computed only in C++ (the Python copies had drifted in their tie-breaking): each GreedyBy allocator just names its GreedyOrder, the Python order_by_* functions stay as thin wrappers over the native orders, and FirstFitPlacer.order reuses a placer's adjacency (genetic seeding at 50k: 2.05 s to 1.32 s). greedy_by_all is omni without linearization, 4-7x faster at 100k and 1M with identical peaks, and its thread pool is gone. Omni's placements on scalar input are byte-identical to the parent.

ConflictSweep owns its reduced rows, happens_before replaces the inverted dominates, one hash_combine replaces three combiners, ClockSpans is move-only since its spans may view its own buffer, the bindings lose their redundant return policies, the raw greedy_order binding checks the total size, and stale comments are corrected. Telamalloc allocates its scratch once and bounds its incumbents by the deadline (160k disjoint allocations with a 1 s timeout took 4.5 s, now 0.26 s), tabu search forbids a reversal for its full tenure, the random allocator gets a timeout and takes (seed, num_trials, timeout) like every other seeded allocator, genetic checks its deadline while seeding, and every seeded allocator validates its seed up front via common.validation.ensure_seed.

On vector clocks only: the contention order draws every key from the timeline it ranks (a linearized surrogate's durations break ties, as DURATION and AREA already do), and greedy_by_all no longer linearizes; genetic's contention seed ranks on the same surrogate. greedy_by_start ranks clock lanes with omni's canonical_starts instead of main's separate Python content sort, so its vector-clock packings follow omni's START order and can differ from main's (39 worse, 26 better of 352 fuzzed instances; scalar unchanged). greedy_by_start breaks ties by input index, as omni does. Stacked on #73.