Honor Pipeline9 bus allowed layers per connection

[ShiboSoftwareDev]

Stacked on #2243.

Real board failure

The regression is extracted from the AM62L carrier HDMI path and keeps the real SII9022ACNU -> TPD12S016PWR -> HDMI_001S three-point connection. TMDS_D2_N belongs to a bus whose allowedLayers is ["top"], but Pipeline 9 previously emitted part of that connection on inner1.

The SVG snapshot in #2243 shows the violating inner-layer segment. This PR updates that same snapshot to show the complete connection on top.

Fix

• derives allowed z-layers for every final MST/point-pair connection from its logical bus member
• carries that constraint on each port point through tiny-hypergraph pathing and high-density routing
• uses a node-wide layer restriction only when every connection in the node has the same layer set
• keeps mixed per-connection constraints separate instead of applying the first port's layers to the whole node
• rejects forbidden tiny-hypergraph port transitions before cost calculation or candidate queueing
• validates the routed HD output against the declared bus layers before returning it

The pre-queue traversal hook is implemented by the standalone, main-based tscircuit/tiny-hypergraph#173. The dependency pin is a temporary compatibility commit built from current tiny-hypergraph main, #173, and the two unmerged solver APIs already consumed by the #2243 base (trace-density options and preferred-route preservation). It excludes the obsolete repro fixture and branch ancestry. This does not add a Pipeline 9 stage and does not add an __allowedZ field to SRJ connections.

Validation

• real AM62L D2-pair regression passes and its SVG snapshot changes from the violation to top-only copper
• ten related Pipeline 9, preloaded-route, and tiny-hypergraph tests pass (11 focused tests total)
• TypeScript typecheck passes
• production build passes
• the compatibility branch's 19 focused tiny-hypergraph tests pass
• tiny-hypergraph's focused transition-filter regression proves rejected neighbors are neither costed nor queued

Known follow-up

The reduced real pair routes correctly. The full eight-member HDMI phase now fails loudly when its current topology requires a single-layer crossing, rather than silently escaping the declared layers. That topology problem is separate from enforcing allowedLayers.

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated (UTC)
capacity-node-autorouter	Ready	Preview	Aug 27, 2026 8:06pm

[tscircuitbot]

Benchmark This PR

Run benchmarks by commenting on this PR:

/benchmark [benchmark.sh args...]
/benchmark-long [benchmark.sh args...]
/benchmark-all [benchmark.sh args except --dataset]
/benchmark --same-machine
/benchmark-all --same-machine
/profile --dataset 18

Comment /benchmark to run the default dataset, or append any arguments accepted by ./benchmark.sh.
Comment /benchmark-long for an 8-vCPU run that defaults to 8 workers and has an eight-hour timeout.
Comment /benchmark-all to start separate workflow runs and result comments for the default dataset plus srj18, srj19, srj20, srj21, and srj23. You may append benchmark arguments such as --pipeline 9; --dataset is not accepted because the command selects every configured dataset.
Append --same-machine to /benchmark or /benchmark-all to compare current main and the PR head sequentially on the same 8-vCPU Blacksmith runner.
Comment /profile --dataset 18 to compare direct Pipeline 7 stage-time percentages between current main and the PR head sequentially on one Blacksmith runner.

Everything after /benchmark, /benchmark-long, or /benchmark-all is safely forwarded to ./benchmark.sh, except --profile-solvers, which enables profile comparison tables, and --same-machine, which enables the paired runner.
Examples: /benchmark --dataset 18 --sample-timeout 2000s, /benchmark --pipeline 10 --dataset 29, /benchmark --pipeline 10 --dataset 29 --sample 1, /benchmark --pipeline 7 --limit 20, /benchmark all 20 --concurrency 8, /benchmark --same-machine, /benchmark-all --pipeline 9 --same-machine, and /benchmark-long --dataset 18.

Use /update-snapshots (or /us) to run BUN_UPDATE_SNAPSHOTS=1 bun test --timeout 120_000 on the PR branch and auto-commit snapshot updates.
Use /usf to read recent failed test files, update and verify their exact CI test shards, and auto-commit only their snapshots. It uses the configured fast benchmark runner by default; use /usf --ubuntu-latest for GitHub-hosted x64 CI parity.

Any PR whose title contains [BENCHMARK TEST] will automatically run one default-dataset benchmark on PR updates; it does not post a PR result comment.

[seveibar]

this doesn't make sense- one connection limits the entire node to particular layers?

also why is __allowedZ being made part of the API?

You're hacking on bus routing to Pipeline9 without a plan

[seveibar]

meant to RC

[ShiboSoftwareDev]

Rewritten from scratch on the real AM62L repro in #2243. The previous first-port/node-wide constraint bug and SRJ __allowedZ approach are gone; constraints now remain per connection, with node-wide restriction only when every connection agrees. The same real-board SVG snapshot is updated to show the fix. Ready for re-review once CI finishes.

[github-actions]

This PR has been automatically marked as stale because it has had no recent activity. It will be closed if no further activity occurs.

[github-actions]

This PR was closed because it has been inactive for 1 day since being marked as stale.