[inureyes]

Reconciles the b10621 compatibility manifest after wave 1 of epic #1431, and corrects a sign error in the teacher-forced trace comparison.

Ownership drift: nine entries filed under an issue that cannot fix them

Each of these carried a divergence naming a different issue than its own issue field, because its shard's owner set left no alternative. They move to the issue that owns closing the gap: GET /health and GET /v1/health to #1440, GET /models and GET /v1/models to #1438, POST /completion and POST /completions to #1466, POST /embedding and POST /embeddings to #1452, and --alias to #1438. pin.json gains the matching owners so the ownership check accepts them.

This does not make any shard editable by two concurrently running chains. routes gains #1438, #1440 and #1466, all of which belong to chain A alongside its existing #1441, #1442 and #1452. model-source gains #1438 next to #1434, and #1434 is closed, so chain B will not edit that shard again.

The reopened issues: twenty-four entries move to tracked follow-ups

Wave 1 merged and closed #1450 and #1453 while twenty-four of their entries were still deferred. The gate requires a deferred entry's issue to be open, so that turned the llama-compat manifest job red for every open pull request in the repository, twice. #1472 now owns the thirteen context and batching entries and #1473 the eleven prompt-cache and slot-scheduling entries, each reproducing the entry's recorded divergence verbatim; runtime-and-context gains both as owners. #1450 remains an owner because six supported entries still reference it.

The issue numbers were shipped to operators

--cache-reuse's startup refusal, the YaRN refusal and the RoPE override refusal all ended with Tracked by #1450. or Tracked by #1453., asserted by their unit tests. Closing those issues would have sent operators to closed issues, so the diagnostics and the assertions now name #1472 and #1473.

The trace comparison reported the reciprocal of perplexity

scripts/compare_logit_traces.py computed exp(-mean(nll)) and printed it under a perplexity header. Column four of a trace is the true NLL and is positive, because examples/logit_trace.rs writes -log_softmax(target), so that expression is the geometric mean probability, which is the reciprocal of perplexity. A candidate that got worse therefore reported a negative percentage under a header that read like an improvement. Eight sub-issues of this epic still owe teacher-forced logit traces, so the sign is load-bearing. Verified against a synthetic pair with NLL 0.1 and 1.0: the fixed script reports 1.1052 against 2.7183 at +145.960%, matching exp(0.1) and exp(1.0) analytically. The line now also says "higher is worse".

Verification

python3 scripts/ci/check_llama_compat_manifest.py passes: 249 help entries, 323 spellings, 136 env vars, 53 routes, 74 native fields, states {supported: 45, aliased: 12, not_applicable: 61, deferred: 258}, 20 distinct deferred issues. A generated cross-check confirms all 17 rows of the shard ownership table in docs/llama-server-compat.md now match pin.json; that check is what caught the chat-templates row still omitting #1470, which wave 1 added to pin.json without updating the table. cargo test --lib cli:: 137 passed, --test llama_compat_manifest 4 passed including the archive-gated conformance test, cargo fmt --all -- --check clean.

Part of #1431.