feat: caller lookups for read-write memory accesses (3/3)

[OlivierBBB]

Part 3 of 3 of the RAM timestamps feature (#1807), stacked on #2080 (which stacks on #2079); the diff shown here is only this PR's changes.

What it does

Every read-write memory access now emits a lookup tying it to a row of that memory's table, through the same per-site machinery as calls and access-once memories: the source tuple carries the accessor's address, data and threaded timestamp registers — gated by the shared line/path selector — and maps onto the table's ADDRESS, VALUE_WRITTEN and TIMESTAMP_WRITTEN columns. Multiple accesses on one row each get their own lookup. A read's data registers are its outputs (the value read back), which the table exposes in VALUE_WRITTEN too, since a read row writes back what it found (VALUE_READ == VALUE_WRITTEN).

The access's read/write kind is pinned by the target-side filter rather than a tuple element (MIR lookup tuples carry register accesses only): the table gains two fully-determined selector columns, EXEC_WRITE = EXEC · IS_WRITE and EXEC_READ (via EXEC_WRITE + EXEC_READ == EXEC), filled by the observer. A write site's lookup targets the table filtered by EXEC_WRITE, a read site's by EXEC_READ.

With this, the guest-facing columns of every access row are pinned by the caller, and the row-local constraints hold end-to-end. The remaining gap — that VALUE_READ genuinely returns the last value written to the address — is the offline memory-checking bus, tracked in #2068 together with the finalization rows.

Testing

The ram_01..10 fixtures and the RAM-heavy benches (blake / sort / keccakf / poseidon) now run with the complete chain live — threading → table → lookups — across the full Test_ZkcUnit|Test_ZkcMixed|Test_ZkcInvalid and bench suites; lint clean.

🤖 Generated with Claude Code