Transitive Taint Propagation for Shared Agent State: An Empirical Evaluation

Version v0.2 of the empirical companion to Transitive Taint Propagation (TTP). This version extends the v0.1 Empirical Evaluation with the measured boundaries of the primitive's bounding levers: (1) the containment denial-of-service surface, with over-containment up to ~92% of the blast radius on a maximally-independent workload under taint-every-read; (2) a read-relevance content-derivation gate that raises blast-radius precision from ~34% to ~100% at zero recall loss when derivation is literally visible, with its paraphrase blind-spot reported as the published recall floor (~98% of true dependencies invisible at the zero-overlap extreme); (3) a no-false-negative (superset) guarantee for summarized containment, scoped to the recorded dependency graph; and (4) a provenance-summary structure that bounds write-side ancestor storage from O(n^2) to O(n) (n^1.00, r^2 = 1.000), trading only a tunable false-positive over-containment. The v0.1 core results (within-session containment recall, the cross-session recall-vs-carry soundness boundary, and the performance envelope) are carried forward so the latest version remains self-complete. Every figure is a seeded, oracle-scored median reported with its operating envelope and honest scope; each lever trades something measured and on the page. Reference-implementation results are reproducible from committed study runners. CC-BY-4.0.