perf: optimize bus balance checking

[OlivierBBB]

Follow-up to #2181. No change to which traces are accepted or rejected — same
verdicts, less work.

Measured, versus main

200,000 selected bus rows, medians of 3, same machine back to back.
alldistinct = every message distinct (the shape RAM consistency will have);
duplicated = 16 distinct messages over all rows.

shape	field	main	this PR
alldistinct	koalabear	30.2 ms	22.8 ms	1.32x
alldistinct	bls12_377	33.1 ms	24.4 ms	1.36x
duplicated	koalabear	13.1 ms	4.50 ms	2.9x
duplicated	bls12_377	20.4 ms	8.63 ms	2.4x

Allocations: 400,556 → 300,560 (alldistinct), 200,041 → 61 (duplicated).
Spread on the alldistinct rows is ~15%, so treat those as approximate.

Measured with a throwaway benchmark, added and then reverted here (commit
6d9eae25 if anyone wants it). Not kept: nothing in the repo runs Go benchmarks
— corset-bench and zkc-bench-test use -run, selecting tests named Bench —
so it would never execute, yet would still be compiled by every go test ./....

What changed

1. Buses no longer checked per shard. Every constraint ran on each shard and
   bus failures were then discarded — including two wasted count() rescans per
   bus, since a lone shard rarely balances. Filters the constraints instead, via
   getBusAndOtherConstraints and the new schema.AcceptsSubset.
2. No per-row allocation. Rows read into a reused buffer. Safe only because
   the new hash.Map.Update retains a key just when it is new, and then clones
   it. Update also folds lookup and store into one visit, halving the hashing.
   This is @DavePearce's review comment on feat: lisp syntax and sharded trace checking for send and receive bus primitives #2181.
3. Port columns resolved once per port, not per row (portColumns).
4. Dropped NetTally (no callers); renamed fixture modules alpha/beta to
   sndr/rcvr; checkShardGroup to checkShards, since a bus balances over
   the complete set of shards, never a subset.

Review comments from #2181

• NewConstraint took its width from the sends alone, so receives-with-no-sends
  panicked with a misleading message. Now taken from whichever side has ports.
  Defensive only: translateBus rejects such a bus first, and Consistent
  checks it again.
• Sharded tests were counted by line, so a comment-only fixture satisfied the
  "missing any tests" guard whilst checking nothing. Counts trace sets now.
• Added bus_invalid_08, covering "receives but no sends" — bus_invalid_04
  covered only the mirror case, leaving that message untested.
• Not fixed here: an indented ;; comment reaches the JSON parser instead of
  being skipped. It affects all three fixture readers rather than the sharded one
  alone, nothing in testdata triggers it, and it is unrelated to bus
  performance. Noted for a separate issue.

Ideas measured and rejected

• Pre-sizing the tally (it starts at 32 entries). Sizing it to 200,000 was
  15-18% slower and used 11x the memory.
• A uint64 fast path for messages narrow enough to pack into one word. Real
  bus messages are (address, value, timestamp) at roughly 64/64/48 bits, and
  ports are built from limbs, so a RAM entry is ~11 registers on a 16-bit
  register field. It would only ever fire on toy fixtures.

Stopping here deliberately

The ~200k allocations still left in the alldistinct case are structural to
hash.Map, which gives every entry two slices. Replacing that — e.g. a Go map
keyed on the message bytes — would cut them, but the tally is likely to be
restructured anyway: a per-shard bus digest for the prover wants it keyed per
shard and per side, neither of which the current global signed tally has.
Further work on its internals has a short shelf life until that is settled.

🤖 Generated with Claude Code

[Copilot]

Pull request overview

This PR optimizes sharded trace checking and bus-balance validation by avoiding per-shard bus evaluation and reducing repeated column lookups during bus scans, while also cleaning up unused bus APIs and updating related fixtures.

Changes:

• Avoids running bus constraints per-shard by splitting constraints into bus vs non-bus once (getBusAndOtherConstraints) and adding schema.AcceptsSubset for checking only non-bus constraints on each shard.
• Speeds up bus constraint/failure rescans by resolving a port’s selector/message columns once per port (portColumns) rather than per row.
• Removes unused NetTally and renames fixture modules (alpha/beta → sndr/rcvr) to match updated test expectations.

Reviewed changes

Copilot reviewed 13 out of 13 changed files in this pull request and generated no comments.

Show a summary per file

File	Description
testdata/corset/valid/bus_05.shards.accepts	Adds spacing between shard-set cases for readability (no semantic change).
testdata/corset/valid/bus_02.rejects	Updates fixture column names to sndr/rcvr.
testdata/corset/valid/bus_02.lisp	Renames fixture modules to sndr/rcvr.
testdata/corset/valid/bus_02.accepts	Updates fixture column names to sndr/rcvr.
testdata/corset/valid/bus_01.shards.rejects	Updates shard fixture column names and adds spacing between cases.
testdata/corset/valid/bus_01.shards.accepts	Updates shard fixture column names and adds spacing between cases.
testdata/corset/valid/bus_01.rejects	Updates fixture column names to sndr/rcvr.
testdata/corset/valid/bus_01.lisp	Renames fixture modules to sndr/rcvr.
testdata/corset/valid/bus_01.accepts	Updates fixture column names to sndr/rcvr.
pkg/test/util/check_legacy.go	Refactors sharded checking to avoid per-shard bus evaluation; introduces constraint partitioning and uses AcceptsSubset.
pkg/schema/schemas.go	Adds AcceptsSubset to run a provided iterator of constraints against a trace.
pkg/schema/constraint/bus/failure.go	Uses cached per-port column lookups when collecting required cells on failure.
pkg/schema/constraint/bus/constraint.go	Removes unused NetTally; adds portColumns and uses it to reduce repeated column lookups in tally/count paths.

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.

[OlivierBBB]

Condensed review comment:

• this PR removes useless shard level checks of buses: checking the bus constraints (its balancing) makes no sense at the individual shard level, and previously this was being done up to 3 times per bus constraint and per shard; the failure message was then discarded and thus completely useless;
• it adds a portColumns function to extract the relevant selector and data columns from a concrete trace using the port blue print; previously when tallying the sends / receives it would "rediscover" the appropriate trace columns at every row of a trace
• fewer allocations with the Update method: update the map busMessage -> count and only allocate a new bus message (key) if it's new from that map's perspective; the key slice you provide the Update method is re-used by the caller many times rather than providing a fresh new slice at every call; this addresses @DavePearce 's message on the previous PR feat: lisp syntax and sharded trace checking for send and receive bus primitives #2181; on an example written by Claude this took 200,041 allocations down to 61
• some name changes in bus tests + ras

[DavePearce]

I am going to close this because it conflicts with #2186, and I don't want this to merge before that one. I took the main pieces from this branch related to memory optimisation and the hash.Map.Update function. However, changes to the test framework are not viable here since I have basically reworked that now for sharding.