Add MEMW_R register fast-path table

[jotabulacios]

Implements the MEMW_R register fast path from spec PR #457.

Adds a new 10-column table for register accesses.
Registers are always aligned 2-word accesses with is_register=1, so this table strips the general-purpose memory columns (address decomposition, alignment mask, width flags, per-byte old timestamps)

• 10 columns, 7 bus interactions, 3 transition constraints.
• Timestamp ordering via IS_HALF lookup instead of LT, proving a delta in [1, 2^16].
• Routing predicate is_register_op sends eligible register ops toMEMW_R at trace-build time.

[github-actions]

Codex Code Review

Findings

1. Low - Security/Availability (panic on malformed input)
   In prover/src/tables/memw_register.rs:113, generate_memw_register_trace uses an unconditional assert_eq! on old_timestamp[0] == old_timestamp[1].
   If logs/witness data are malformed (or upstream routing regresses), this can crash the prover process (DoS) instead of returning an error.
   Action: replace with fallible handling (Result) or explicitly reroute/reject without panicking.

2. Low - Potential bug (release-only invariant gap)
   In prover/src/tables/memw_register.rs:103, register-address parity is guarded only by debug_assert_eq!, but the code still does base_address / 2 in release mode.
   If an odd base_address slips through, it is silently truncated and can cause confusing downstream proof failures rather than a clear error at source.
   Action: enforce this invariant in release builds too (explicit check + error), not just debug builds.

No other concrete security/correctness/performance issues stood out in the PR diff.

[jotabulacios]

/bench

[github-actions]

Benchmark — fib_iterative_8M (median of 3)

_{Table parallelism: 32 (auto = cores / 3)}

Metric	main	PR	Δ
Peak heap	250294 MB	220983 MB	-29311 MB (-11.7%) 🟢
Prove time	79.886s	47.095s	-32.791s (-41.0%) 🟢

🎉 Improvement detected — heap or time decreased by more than 5%.

✅ Low variance (time: 2.5%, heap: 0.2%)

_{Commit: b1f365b · Baseline: cached · Runner: self-hosted bench}

[github-actions]

Codex Code Review

Findings

1. Medium (Security/Robustness): panic on malformed register timestamp state

• In memw_register.rs:113, generate_memw_register_trace uses assert_eq!(op.old_timestamp[0], op.old_timestamp[1], ...).
• This is in non-test production code and will panic! instead of returning a recoverable error if the invariant is violated. In prover services, this is a crash/DoS vector.
• Action: replace with explicit validation that returns Result<..., Error> (preferred), or at minimum downgrade to debug_assert_eq! and enforce routing invariants before this function is called.

2. Low (Potential bug in test coverage): vacuous LT assertion

• In trace_builder_tests.rs:495, the test asserts total_lt_rows > 0.
• This can pass even when there are no real LT ops, because empty LT op sets still generate padded trace rows.
• Action: assert on meaningful LT multiplicity or a specific non-padding LT condition, not row count.

Could not run tests in this environment due sandboxed read-only ~/.cargo/~/.rustup paths blocking dependency/toolchain setup.

[claude]

Review: MEMW_R register fast-path table

Overall the design is sound and well-structured. The structural sharing of TIMESTAMP_1 as the high limb of both timestamps is a clever way to avoid an extra constraint, and the IS_HALFWORD substitution for LT is valid given the routing predicate enforces matching upper limbs. Three issues noted via inline comments:

Severity	Issue
Medium	collect_bitwise_from_memw_register: underflow in ts_lo - old_ts_lo - 1 is only safe because of implicit pre-filtering; silent wrong result in release mode if invariant breaks
Medium	assert_eq! in generate_memw_register_trace panics in release builds; should be debug_assert_eq! since the routing predicate already guards this
Low	constraints() is missing a comment explaining why no old_ts_hi == ts_hi constraint is needed (structural column sharing is the reason, but it's a non-obvious load-bearing invariant)

[diegokingston]

Code review

No issues found. Checked for bugs and CLAUDE.md compliance.

🤖 Generated with Claude Code

[jotabulacios]

/bench