docs(qwen3): remove stale docs and dead code

[Nyvo-io]

Summary

• remove the unused ModelInfo type left in openinfer-engine after the Qwen3 model-probe path was retired
• replace stale Qwen3 crate and tensor-parallel bring-up records with current ownership/runtime documentation
• move the durable KV pressure lesson into docs/lessons/, delete obsolete model records, and repair indexes/cross-references
• record landed real-adapter LoRA and cached-token propagation gates without presenting them as open work

Maintainer Follow-up

Commit 656da43 addresses the requested roadmap correction against the current base:

• openinfer-qwen3/tests/lora_golden_gate.rs covers a non-zero PEFT adapter for bs=1, mixed base/LoRA batches, TP=2 sharding, and a silently inactive adapter guard
• openinfer-qwen3/tests/cached_tokens_usage.rs and openinfer-vllm-frontend/src/bridge/tests.rs cover TokenEvent::Scheduled.cached_tokens propagation through first-output and terminal-output frontend paths
• the remaining roadmap gaps are TP numerical parity, aggregate prefix-cache telemetry, LoRA cache-salt isolation, and YaRN [Feature Request]: Support YaRN (Yet another RoPE extensioN) for context window extension #8

The rebase also accounts for changes that landed independently on main:

• chore(tooling): add hawk config + playbook; drop 131 dead_public items #743 already removed the Qwen3 probe_model path and dead trace constants, so this PR no longer duplicates those removals; the current diff removes only the residual unused ModelInfo
• refactor(qwen35): use size-neutral crate and feature names #747 renamed the Qwen3.5 crate to openinfer-qwen35; the conflict resolution keeps the new path while preserving the allocator-specific KV distinction
• Qwen3 commits at most P + N - 1 positions, but kvbm can provision a dangling decode block, so a multi-token request reserves ceil((P + N) / block_size) at peak; Qwen3.5 retains its allocator-derived P + N - 1 formula
• stale references to the removed kernel_plan and the retired openinfer-core sampling path were replaced with the current DAG, manifest/report, and openinfer-kernels surfaces

Verification

• base: 94975a1 (origin/main, including refactor(qwen35): use size-neutral crate and feature names #747)
• head: 656da43
• scope: 10 files, two Nyvo-authored/Nyvo-committed DCO-signed commits
• GitHub CI run 30002273553 and attribution run 30002272111: all 15 reported checks pass, including Rust formatting, DCO, attribution, locked Cargo metadata, CPU Clippy/tests, simulated frontend E2E, Qwen3 CUDA compile, and Qwen3 CUDA Clippy (sm_80)
• two post-rebase full-diff reviews found no remaining behavioral regression or blocking documentation error
• deleted paths/symbols and stale crate/API references have no remaining references in the affected surface; git diff --check passes and the worktree is clean

Closes #248

[Nyvo-io]

Status update: PR #702 remains ready for maintainer review. The current head 8b0ad956 is clean, all reported CI checks are passing, and there are no new review comments or requested code changes. I have kept the scope limited to the Qwen3 documentation/dead-code cleanup and the durable KV-pressure lesson.

[Nyvo-io]

Final update after rebasing PR #702 onto origin/main@afd7a64:

• final head is e1f9fb5, a single Nyvo-authored, Nyvo-committed, DCO-signed commit
• the only rebase conflict was the openinfer-qwen3/src/lib.rs import block; the resolution preserves the current upstream rustfmt/import layout and all new runtime modules while continuing to remove only the unused ModelInfo/probe_model path
• the 12-file scope is unchanged, deleted symbols and document paths have no remaining repository references, and the PR is clean/mergeable
• final CI run 29857427251 passes all 15 checks, including formatting, DCO, attribution, CPU Clippy/tests, simulated frontend E2E, Qwen3 CUDA compile, and Qwen3 CUDA Clippy

Please review when convenient.

[xiaguan]

This is not fully addressed yet: the refreshed roadmap still lists the real-adapter LoRA gate and cached-token propagation as open, although both already exist with regression coverage on this PR's base.

[Nyvo-io]

Final update after rebasing PR #702 onto origin/main@94975a1:

• final head is 656da43; both commits are Nyvo-authored, Nyvo-committed, and DCO-signed
• the maintainer-requested roadmap correction is complete: the real-adapter LoRA gate and cached-token propagation are recorded as landed, with their actual regression coverage; remaining work is limited to TP numerical parity, aggregate cache telemetry, LoRA cache-salt isolation, and YaRN [Feature Request]: Support YaRN (Yet another RoPE extensioN) for context window extension #8
• chore(tooling): add hawk config + playbook; drop 131 dead_public items #743 had already removed the Qwen3 probe/trace dead code, so the current 10-file diff removes only the residual unused ModelInfo and does not duplicate upstream work
• the final review pass also corrected the allocator-specific KV peak distinction, removed a stale kernel_plan claim, and replaced the retired sampling source path
• two post-rebase full-diff reviews found no remaining behavioral regression or blocking documentation issue; deleted symbols/paths and affected stale references scan clean, git diff --check passes, and the worktree is clean
• CI run 30002273553 plus attribution run 30002272111 pass all 15 checks, including Qwen3 CUDA compile and CUDA Clippy
• GitHub reports the PR open, non-draft, CLEAN, and MERGEABLE; CAICAIIs remains requested for review

Please re-review when convenient.

[github-actions]

This pull request has been inactive for 14 days. It will be closed after another 30 days unless there is new activity.