Add explicit Fastly delivery-compression support #362

Description

Add reusable delivery-compression intent/capabilities and Fastly finalization so an application can explicitly hand off identity bytes for final client compression. Priority remains P3. The paired framework PR #452 and trusted-server#884 drafts contain the extraction, but neither branch-level implementation nor local guest output establishes mainline delivery or live Fastly negotiation.

Steps to reproduce

Inspection steps (no deployment, compression measurement or fresh tests executed):

1. Inspect main Fastly response conversion; main has no delivery-compression intent/finalizer integration.
2. Compare the open framework draft with the consumer draft.
3. Read their exact-revision evidence and limitations, including the reported local debug/WASM loader discrepancy and still-unverified real Fastly negotiation.

Expected behavior

An explicit, checked identity-body intent and unavailable-by-default adapter capability let callers choose offload before omitting their encoder. Fastly applies final headers to buffered and custom streaming sends without collecting lazy bodies; unmarked and unsupported paths retain existing behavior.

Impact

Applications currently duplicate generic delivery intent/capability/header logic, and standard conversion alone misses consumers that send through a custom lazy streaming path.

Proposed fix

Complete and merge the core intent/capability API and Fastly finalizer in PR #452, then move Trusted Server to a merged release and remove its duplicate generic glue. Finalization must follow late header/privacy effects, set X-Compress-Hint: on, maintain repeated/wildcard Vary: Accept-Encoding correctly and prevent stale encoding/length framing.

Applications must provide identity bytes; removing Content-Encoding from compressed bytes is invalid. Preserve gzip/Brotli limitations and explicit unsupported-adapter fallback. Keep negotiation policy, SSAT eligibility, codecs, auctions, HTML/template behavior and privacy application-owned; do not change origin negotiation or add a publisher flag.

The drafts record framework 157f091e42b5c584898a5d0f8a0d3ddc3a4dd5ec and consumer 85a9b9184a67a9c827be9b59ead95a801fb07faf. Their test/CI records are historical branch evidence; refresh required checks before merge. The smoke-fixture EOF handshake fix does not establish live negotiation or performance.

Done when

• Merge and document portable intent/capability APIs with checked identity-body and explicit unsupported-adapter contracts.
• Apply the Fastly contract at final send in buffered and custom streaming paths, after late mutations, without whole-body collection.
• Cover marked/unmarked responses, unsupported adapters, framing/body agreement and repeated/wildcard Vary.
• Retain current behavior for unmarked and already-encoded responses.
• Integrate a merged EdgeZero release into Trusted Server, remove duplicate generic glue and preserve its negotiation/fallback, SSAT, templates and privacy tests.
• Verify real Fastly negotiation separately from local/Viceroy guest output.
• Resolve or document the reported local loader discrepancy and retain exact-revision evidence; do not claim wire-size/vCPU gains without fresh platform measurements.

Related

PR #452 is the fully covering open framework PR; trusted-server#858, trusted-server#884 and original trusted-server#853.