[Sad Catapella B5][Performance] Repeated linear manifest lookups make unchanged runs quadratic in file count #91

Baseline / classification

Adversarial audit Batch 5 at d7a538e56d41ac1093ba79adcad397c21b3dfd32. Confirmed asymptotic inefficiency, runtime slowdown unmeasured.

Evidence

• RunManifest.file() lines 296–298 linearly searches self.files for each relative path.
• run_experiment() lines 524–543 calls previous.file(rel) for every file, making the lookup portion O(N²) for N unchanged images.
• check_sources() lines 518–534 repeats this for each supplied path, leading to O(M×N) lookup work for M caller images. The code also creates a redundant sorted list for manifest leaves; see RunManifest.leaves().
• Tests cover only three entries. Reuse tests cover small experiments here.

Fix / benchmarks

Create an immutable path→FileEntry lookup once (or explicitly build one per batch from the manifest) and use it consistently in run and source checks, with duplicate-path validation. Preserve existing RunManifest JSON schema/order. Benchmark 100, 1k, 10k, 100k images in a no-change rerun and check_sources; compare slope/latency and memory overhead. The iterator/tuple public interface can remain unchanged. Numba is not suitable: Python dictionary indexing removes the algorithmic problem.