[Post-0.31][Backend] Evaluate policy dispatch alternatives with measured crossover evidence #37

Activity

1. a19q3

   added this to the 0.32 milestone on Oct 3, 2026
2. a19q3 commented on Oct 4, 2026

   

   a19q3

   on Oct 4, 2026

   CollaboratorAuthor

   More actions

   Completed the bounded 0.32 dispatch evaluation. Retain the existing linear dispatcher and direct action stubs. No production crossover threshold or new optimization profile is admitted.

   • Balanced routing: 402 paired cases across dense/sparse/high-u32 tags and sizes 1–64; 314 byte/cycle tradeoff rows, 82 favoring linear routing and 6 ties.
   • Saved selector: 804 cases; 205 tradeoff rows and 599 favoring the current design. The two-pass probe includes a clobbering common-check surrogate and distinguishes unknown-tag rejection from common-check failure.
   • Late-action savings do not remove early-path regressions/code growth. The routing microbenchmarks embed the selected input and support only their finite local comparisons; they are not a production threshold or whole-transaction speedup claim.
   • The retained production matrix and independent checker continue to enforce unknown-tag prechecks, common-check/first-failure order, exact decoder/action destinations, selector dataflow and outgoing-argument ownership. No new machine shape is adopted, so the current versioned checker contract remains authoritative.

   See the experiment decisions, paired frontier and historical full-gate record. CI/backend passed at clean source 6a453d3015784784a27ba1aa0a4b6b1a73db79a8; unchanged ceilings and 6,782 comparable metrics passed. Archived source identities are not relabeled as the latest client commit.

   Closure audit on 2026-10-04: all seven archive byte counts/SHA-256 values in that gate record match. Fresh focused replay on the 0.32 checkout based on 35bf983d passed 9 policy-artifact-checker tests and 6 cost_experiments tests, including both dispatch alternatives. Commands: cargo test --locked -p cellscript --test policy_artifact_checker -- --test-threads=1 (run with cost_measurement in the same invocation), and cargo test --locked -p cellscript cost_experiments --lib -- --test-threads=1.

   This resolves the issue's explicit option to retain current routing when measurements do not justify replacement. A jump table or broader routing policy still needs its own design/evidence; none is implied by closure.
3. a19q3

   closed this as completedon Oct 4, 2026