Design spec for streaming HTTP responses through the publisher proxy
when Next.js is disabled. Covers two implementation steps:

1. Make the streaming pipeline chunk-emitting (HtmlRewriterAdapter,
   gzip, encoder finalization)
2. Wire up Fastly StreamingBody via stream_to_client() with entry
   point migration from #[fastly::main] to undecorated main()

Includes streaming gate logic, error handling, rollback strategy,
and testing plan. Verified against Fastly SDK 0.11.12 API.

…er, rollback

…reaming refactor

…ssing paths

…ation test

- Add debug-level logging to process_chunks showing total bytes
  read and written per pipeline invocation
- Add brotli-to-brotli round-trip test to cover the into_inner()
  finalization path
- Add test proving HtmlWithPostProcessing accumulates output when
  post-processors are registered while streaming path passes through

- Group std imports together (cell, io, rc) before external crates
- Document supported compression combinations on PipelineConfig
- Remove dead-weight byte counters from process_chunks hot loop
- Fix stale comment referencing removed process_through_compression
- Fix brotli finalization: use drop(encoder) instead of into_inner()
  to make the intent clear (CompressorWriter writes trailer on drop)
- Document reset() as no-op on HtmlRewriterAdapter (single-use)
- Add brotli round-trip test covering into_inner finalization path
- Add gzip HTML rewriter pipeline test (compressed round-trip with
  lol_html, not just StreamingReplacer)
- Add HtmlWithPostProcessing accumulation vs streaming behavior test

- Add Eq derive to Compression enum (all unit variants, trivially correct)
- Brotli finalization: use into_inner() instead of drop() to skip
  redundant flush and make finalization explicit
- Document process_chunks flush semantics: callers must still call
  encoder-specific finalization after this method returns
- Warn when HtmlRewriterAdapter receives data after finalization
  (rewriter already consumed, data would be silently lost)
- Make HtmlWithPostProcessing::reset() a true no-op matching its doc
  (clearing auxiliary state without resetting rewriter is inconsistent)
- Document extra copying overhead on post-processor path vs streaming
- Assert output content in reset-then-finalize test (was discarded)
- Relax per-chunk emission test to not depend on lol_html internal
  buffering behavior — assert concatenated correctness + at least one
  intermediate chunk emitted

lol_html fragments text nodes across chunk boundaries when fed
incrementally. This breaks script rewriters (NextJS __NEXT_DATA__,
GTM) that expect complete text content — a split domain like
"google" + "tagmanager.com" would silently miss the rewrite.

Add dual-mode HtmlRewriterAdapter:
- new(): streaming mode, emits output per chunk (no script rewriters)
- new_buffered(): accumulates input, feeds lol_html in one write()
  call on is_last (script rewriters registered)

create_html_processor selects the mode based on whether
script_rewriters is non-empty. This preserves the old behavior
(single-pass processing) when rewriters need it, while enabling
streaming when they don't.

Add regression test proving lol_html does fragment text across
chunk boundaries, confirming the issue is real.

lol_html fragments text nodes across input chunk boundaries. Script
rewriters (NextJS __NEXT_DATA__, GTM) expect complete text content
and would silently miss rewrites on split fragments.

Add dual-mode HtmlRewriterAdapter:
- new(): streaming, emits output per chunk (no script rewriters)
- new_buffered(): accumulates input, single write() on is_last

create_html_processor selects mode based on script_rewriters. This
preserves correctness while enabling streaming for configs without
script rewriters. Phase 3 will make rewriters fragment-safe.

Add regression test proving lol_html does fragment text nodes.

Add section to spec explaining the lol_html text fragmentation issue,
the dual-mode HtmlRewriterAdapter workaround (Phase 1), and the
planned fix to make script rewriters fragment-safe (Phase 3, #584).

- Add post-finalization warning to buffered path (was only in streaming)
- Add buffered_adapter_prevents_text_fragmentation test proving
  new_buffered() delivers complete text to lol_html handlers
- Fix spec: html_processor.rs is changed (selects adapter mode), and
  script_rewriters do require buffered mode (not "unaffected")

Both review comments apply to Phase 2 as a whole, not individual steps.
Move the EC implementation note to the Phase 2 header in the plan and
the Step 2 header in the spec.

…eaming-pipeline-phase1