Harden agent publish, public STDF evidence, lints, fuzz, and decode throughput - #3

Summary

Five reviewable commits for the top priorities in the Lauren Tan-derived improvement plan. The stale crates/ tree is untouched. Independent review is pending: this was one sequential agent, and the author is not an independent reviewer of the decoder change in commit 4.

1. Stop hook no longer publishes. .cursor/hooks/auto-commit-push.ps1 writes a local patch under target/agent-work/checkpoints/ and does not git add, commit, or push. ZSTDF_AGENT_PUBLISH is ignored. scripts/publish_agent_branch.ps1 is the explicit coordinator path: it refuses main/master, runs fmt, clippy, and tests, and pushes only HEAD:refs/heads/agent/.... scripts/check_pr_size.py warns (does not fail) above 800 non-generated lines.
2. Pinned public STDF counts are a CI assertion. examples/public-stdf/expected.json holds the decoder and viewer counts. The test checks SHA-256 from sources.json and skips unless the files exist or ZSTDF_REQUIRE_PUBLIC=1. Measured counts matched the README, including a595 latest 270 / unresolved 66.
3. Workspace lints, toolchain pin, cargo-deny. unsafe_code is forbid, except stdf-io mmap and stdf-parquet FFI, which deny it and allow it only on those functions with SAFETY comments. unwrap_used and panic are denied outside tests. Production sites were turned into Result or a non-panicking exit. rust-toolchain.toml pins 1.99.1. deny.toml ignores advisories that parquet 55 / pyo3 0.24 cannot upgrade past; the reasons are in the file.
4. Property tests and cargo-fuzz. stdf-core proptest covers a minimal FAR/PIR/PTR/PRR/DTR/MRR/unknown round trip, entry-point agreement, and prefix stability. fuzz/ has decode_all and truncate_prefix targets using the same oracles, seeded from the golden hex fixtures. A libFuzzer smoke found decode_all accepting a non-v4 FAR that decode_with_summary rejected. Both decode_all and summarize now use detect_byte_order_result.
5. Decode-throughput experiment. Criterion benches the pinned files when they are downloaded, plus a synthetic file. scripts/bench_adapter.py prints {"records_per_sec": ...}. experiments/decode-throughput-01/contract.json protects tests, golden fixtures, expected.json, and the adapter. min_relative_gain is 0.30, above a 10-run max deviation of 0.141 and a later same-binary sample about 0.20 above that median.

Verification

• python3 -m unittest scripts/test_check_pr_size.py — 6 passed. scripts/test_improvement_runner.py included in a 40-test run, passed.
• PowerShell 7.4.6: hook on a temp repo wrote a checkpoint, left HEAD unchanged, and did not copy untracked bytes. publish_agent_branch.ps1 refused main, agent/../evil, and agent/Main before cargo.
• Public files downloaded and checksums matched sources.json. cargo test -p stdf-cli pinned_public_files_match_expected_counts with ZSTDF_REQUIRE_PUBLIC=1 passed (about 107s).
• cargo clippy --workspace --locked -- -D warnings passed after the lint fixes.
• cargo deny --locked check (cargo-deny 0.20.2): advisories, bans, licenses, and sources ok. Duplicate-version warnings remain warnings.
• cargo test --workspace --locked --exclude stdf-py with ZSTDF_REQUIRE_PUBLIC=1 passed, including after the FAR alignment (stdf-cli 189 tests).
• cargo test -p stdf-core --test properties passed (round trip, arbitrary bytes, prefixes, golden fixtures).
• libFuzzer, nightly 1.101.0, 15s each after the FAR fix: decode_all 5550 runs, truncate_prefix 1548 runs, both exit 0. The first smoke crashed on the FAR mismatch above; that input is fixed, not ignored.
• Criterion --sample-size 10: synthetic plus lot2, lot3, and a595 completed.
• scripts/bench_adapter.py printed one JSON object. improvement_runner.py init on the new contract returned ready.

Not run / follow-up

• cargo test -p stdf-py does not link here. pyo3 is built with extension-module, so the test harness has no libpython. python3-dev was installed and the link line still omitted -lpython. This matches docs/development/multi-agent.md (cargo test --workspace --exclude stdf-py locally). CI still runs the full workspace test on Windows.
• cargo fmt --all -- --check is not clean on rustfmt 1.10.0-stable (1.99.1). That was already true before this branch; this change did not reformat the tree. New and edited Rust files were formatted.
• cargo clippy --all-targets -- -D warnings fails on pre-existing clippy::approx_constant in unit tests (3.14). Default CI clippy does not pass --all-targets.
• expect_used and indexing_slicing were not enabled. CI uses -D warnings, so a warn-level lint would fail the build, and indexing_slicing would flag ordinary parser indexing.
• The full improvement-runner evaluate loop was not executed. Only init and the noise samples were.
• Local libFuzzer needed g++ and LIBRARY_PATH to libstdc++. GitHub's Ubuntu image usually has that toolchain; if the new fuzz job cannot find libstdc++, set the link path there.
• Independent review of the decoder FAR change is pending.