[SunSi12138]

Summary

Implements #550 against dev, following the latest owner architecture addenda.

• Adds direction-specific instance-scoped runtime send-policy APIs for Client Request compression and Server Response compression.
• Keeps SharpLinkCompressionOptions capability-only (Providers / wire profiles); initial Request/Response send policies live independently in Client/Server build composition.
• Publishes each local send policy as one immutable reference snapshot; each outbound business message performs one snapshot read and uses that snapshot for the full compression decision.
• Keeps provider/profile negotiation immutable and does not add a CompressionPolicyControl capability.
• Carries the initial Client response preference in the handshake as (ResponseCompressionPreferenceGeneration, AllowResponseCompression).
• Uses narrow ResponseCompressionPreferenceUpdate / ResponseCompressionPreferenceAck control frames for later changes.
• Uses cumulative ACK convergence semantics (AppliedGeneration >= requested) with per-session O(1) in-flight/latest coalescing.
• Uses a fixed convergence cohort for SetResponseCompressionPreferenceAsync; propagation is non-fail-fast across captured sessions, closed sessions leave deterministically, and failures are reported only after the captured cohort has been attempted.
• Caller cancellation cancels only that convergence wait and is propagated as cancellation even for multi-session cohorts; the already-published desired preference is not rolled back.
• Reconciles the latest desired response preference after actual Ready publication for fixed/static/dynamic topologies, closing the handshake -> Ready publication race.
• Future/reconnected sessions inherit the latest desired response preference but are not added to an already-running convergence cohort.
• Does not create remote preference-control state when Compression was not actually negotiated / no binding exists.
• Keeps inbound decompression independent of local outbound Enabled/threshold policy and treats ACK as convergence, not a wire barrier.

Targeted acceptance evidence

• Deterministic handshake -> Ready publication race coverage for fixed, static, and dynamic topologies; no sleep-based race assertion.
• Public dynamic Client/Server policy behavior across unary, client-streaming, server-streaming, and duplex RPC shapes.
• Build-time directional Enabled = false with providers still negotiated, followed by runtime enable on the same live connection.
• Reconnect coverage proving a new session inherits the latest Client response-compression preference without another Set call.
• Fixed-cohort non-fail-fast coverage using a real backpressured RpcSession send queue: an earlier ResourceExhausted does not prevent a later healthy session from receiving and ACKing the update.
• Multi-session caller-cancellation regression coverage: both negotiated sessions receive the update, caller cancellation remains OperationCanceledException, and never turns into an AggregateException.
• Capability-threshold ownership guard: no SharpLinkCompressionOptions.Minimum* runtime authority remains.
• Compression runtime documentation now presents the public Capability / RequestPolicy / ResponsePolicy / ResponsePreference model, build-time and runtime policy APIs, fixed-cohort/ACK semantics, reconnect inheritance, and caller-cancellation behavior.

Validation

Validated as one clean implementation commit on the current dev lineage (39f980c268632f7469bf2a9c5601e212c9a02c9e on c324ae237f81d45f8df2e0734c312d91ba002fed) using an isolated GitHub Actions final gate:

• formatting verification
• production project-reference boundary gate
• maintainability gate and gate tests
• Release build: 0 warnings / 0 errors
• Unit tests: 1505 passed
• Generator tests: 264 passed
• Load-test tests: 60 passed
• Integration tests: 428 passed (serial)
• NativeAOT transport/topology smoke, including TCP/shared-memory, SharpPack sidecar, and pre-credit coverage
• pack
• package metadata/contracts verification: 8 package + symbol pairs
• package smoke / reference-rooting smoke
• final single-commit lineage, clean-worktree, and workflow-leak checks
• final dev stability check before publication

Dynamic compression hot-path evidence

Same unary byte[] RPC shape over loopback TCP, median of five measured rounds after warmup:

State	Allocated bytes/op	Provider attempts/op
no-provider	2920.78	0
negotiated-disabled	2920.78	0
enabled	2920.66	1
after-update-disabled	2920.78	0

The disabled negotiated path showed no measured allocation increase versus the no-provider path, while the semantic attempt counters prove that only the enabled state reaches the compression provider.

Pull-request CI

Current-head PR checks are green:

• PR Fast
• CodeQL
• Codec Padding Security Evidence

The first review pass is fully resolved. The second review follow-up fixes caller-cancellation propagation and completes the compression runtime documentation; the updated head has passed the full isolated gate and current-head PR CI.

Closes #550

[SunSi12138]

整体方向基本对：instance-scoped local send-policy snapshot、窄化的 Response preference wire intent、handshake 初值、cumulative ACK、O(1) in-flight/latest，以及 compression hot path 在 rent/provider 前完成 binding/policy/preference/size 判断，这些都符合 #550 最终收敛目标。

但我认为现在还不能把 #550 视为完整收口，原因不只在 correctness：

1. inline 有一个真实的 handshake -> Ready publication race，会让新 session 永久停在旧 response preference；这个需要修并补 deterministic race coverage。
2. inline 的 build/runtime model 还少收了一层：runtime authority 已经拆开，但 SharpLinkCompressionOptions 仍长期承载 threshold，并由它创建初始 send policy；这和 #550 最终 addendum 里“CompressionOptions 只剩 Providers / directional policy 独立”的模型不一致，也没有干净的 build-time Enabled=false directional initial policy。
3. inline 的 cohort propagation 对第一个 live send failure 是 fail-fast；desired state 已经 publish，但后面的 cohort session 连 reconcile 都不会执行，生产环境一次 send-queue congestion 会留下额外的 stale Ready sessions。
4. 验收证据还不够。当前 Fast / Extended / CodeQL 都是 green，但 Extended 只有 load smoke，没有 #550 明确要求的 no-provider / negotiated-disabled / enabled / after-update hot-path A/B；PR 里新增测试也集中在 protocol/session 单元层，没有覆盖 public Client/Server API 的 fixed cohort、reconnect/Ready-publication race，以及 Request/Response/ClientStreaming/ServerStreaming/DuplexStreaming 的 dynamic policy integration。这个 PR 改的是每个 outbound business frame 的路径，建议合并前把性能和 lifecycle/integration evidence 补齐，而不是用全量测试通过代替针对性验收。

除了这些点，provider/profile immutable、inbound decode 独立、ACK 非 wire barrier、stale generation 不回滚等核心状态机目前看是合理的，不建议再扩大抽象。

[SunSi12138]

Review follow-up is complete on current head 6ccc33dd932786c0c4bba5a7df314fd7bdbb245e (single commit on current dev).

Final acceptance evidence:

• Release build: 0 warnings / 0 errors
• Unit: 1504/1504
• Generator: 264/264
• Load-test: 60/60
• Integration: 428/428, serial
• deterministic Ready-publication race coverage for fixed/static/dynamic topologies
• public dynamic policy coverage for unary, client-streaming, server-streaming and duplex
• build-time Enabled=false + same-connection runtime enable
• reconnect inherits the latest response-compression preference
• non-fail-fast fixed-cohort propagation under a real backpressured send queue
• NativeAOT, pack, package verification and package smoke all pass
• hot-path evidence: no-provider / negotiated-disabled / after-update-disabled measured the same ~2920.78 B/op, with provider attempts 0 / 0 / 0; enabled reaches the provider at 1 attempt/op

Current-head PR Fast, CodeQL and Codec Padding Security Evidence are all green. All review threads are resolved. Ready for review.

[SunSi12138]

复审当前 head 后，第一轮三个问题都确认闭合：Ready publication race 已用 publication 后 reconcile + deterministic 三 topology test 修掉；capability / directional policy 已真正拆开；fixed cohort propagation 不再因前部 live failure 放弃后续 session。四种 RPC shape、reconnect、build-time disabled/same-session enable，以及当前 hot-path evidence 也已经补齐到足够支持这个架构方向。

现在只剩两个值得收口的点，见 inline。第一个是 public cancellation 语义的实际 bug，建议修后补一个 2-session cancellation test；第二个是交付/可用性问题：旧 threshold API 已删除，但用户文档还没有告诉调用方新的 directional builder/runtime API 和 response preference 怎么配置。

性能方面，runner 已经实际记录 allocation / CPU/op / elapsed/op / QPS，而且 no-provider 分支在代码上仍然先于 policy snapshot load；我不建议为了这一处 pointer-load/predictable-branch 再扩大实现。P50/P99 若最终 evidence 流程已有统一输出可继续记录，但我不把它作为这轮代码 blocker。

[SunSi12138]

最终复审当前 head 39f980c268632f7469bf2a9c5601e212c9a02c9e：没有发现新的 blocking issue，可以合并。

确认：

• Ready-publication race、capability/policy ownership、fixed-cohort fail-fast 三个首轮问题均已闭合；
• caller cancellation 已恢复标准 OperationCanceledException 语义，并有 2-session regression；
• runtime compression 文档已补齐 Capability / RequestPolicy / ResponsePolicy / ResponsePreference 四层模型；
• local policy 继续是 instance-shared immutable snapshot，per-message hot path 无 session fan-out / lock / allocation；no-provider 在 policy snapshot load 前返回；
• remote preference 保持 cumulative ACK + one in-flight/latest pending 的 O(1) bounded state；
• fixed/static/dynamic Ready-gap、reconnect、build-time disabled -> same-session enable、unary/client-streaming/server-streaming/duplex、backpressured cohort failure 等 targeted coverage 已具备；
• 当前 head PR Fast / CodeQL / Codec Padding Security Evidence 均 green，PR mergeable。

P50/P99 如后续统一 performance evidence 流程需要可继续记录，但按前一轮 review 结论不作为本 PR 的代码 blocker。当前实现与 #550 最终架构目标一致。