Fix response connection persistence handling - #788

Problem

HTTP/1.1 connections are persistent unless either peer requests closure. Fluxzy instead gives a response without an explicit Keep-Alive: timeout an implicit one-second protocol idle timeout. The pool subtracts 200 ms from that value, so a healthy connection can be discarded after roughly 800 ms even though the origin never limited its lifetime.

Changing that default without considering HTTP/1.0 would make HTTP/1.0 responses persistent even when they did not negotiate Connection: keep-alive. Connection options are also comma-separated token lists, but the existing whole-value comparison misses values such as Connection: custom, close and Connection: custom, keep-alive.

HTTP/1.0 persistence is safe only when the response has a self-defined message length. Treating an unbounded HTTP/1.0 body as persistent can leave the proxy waiting for EOF while the origin waits for the next request. A non-final-chunked Transfer-Encoding is likewise close-delimited and overrides any accompanying Content-Length; trusting that length can recycle a connection with unread transfer-coded bytes. A bare 205 Reset Content is not self-delimiting without explicit zero-length framing.

Solution

Represent an absent HTTP/1.1 protocol idle limit as -1, matching the pool's existing "no peer-supplied limit" sentinel. Preserve explicit Keep-Alive timeout/max handling and the existing global unused-pool timeout.

Retain the source response version while parsing flat HTTP/1 headers. HTTP/1.0 remains close-by-default and is reusable only when Connection: keep-alive is present and the response is self-delimiting through Content-Length or valid no-body status semantics. Direct field-based responses are H2 and do not inherit HTTP/1.0 behavior.

Parse Connection as a comma-separated, case-insensitive token list with optional whitespace. Use that parser consistently for close, keep-alive, and upgrade decisions.

For a response whose final transfer coding is not chunked, remove Content-Length, reset the parsed length, and force close-delimited handling for both HTTP/1.0 and HTTP/1.1.

Benefits

• Preserves ordinary HTTP/1.1 connections until the peer or configured global idle policy closes them.
• Keeps HTTP/1.0 close-by-default and prevents indefinite reads of non-self-delimited HTTP/1.0 bodies.
• Prevents connection desynchronization when non-final-chunked transfer coding accompanies a misleading content length.
• Honors compound and repeated Connection fields without substring matching.
• Retains explicit timeout, maximum-request, close, and protocol-upgrade behavior.
• Keeps H2/direct-header response semantics unchanged.

The original current-master performance campaign used four request waves separated by 1.25 seconds:

New measured upstream TLS connections fell from 64/61/224 to zero. Across all three workloads p99 fell 58-65%, allocation/request fell 53-69%, and CPU/request fell 34-38%. The later HTTP/1.0, token-parsing, and transfer-framing hardening does not alter the measured HTTP/1.1 path, but was not separately benchmarked.

Changes

• Default ResponseHeader.TimeoutIdleSeconds to the existing unlimited sentinel.
• Detect HTTP/1.0 response status lines and require explicitly negotiated, self-delimited persistence.
• Add allocation-free token parsing for comma-separated header values.
• Apply token parsing to response close, keep-alive, and upgrade options.
• Normalize non-final-chunked transfer coding as close-delimited and discard conflicting content length.
• Cover implicit and explicit HTTP/1.1 persistence, HTTP/1.0 framing, compound/repeated connection options, H2 construction, transfer-coding precedence, status 205, and real TLS connection reuse.
• Allow the in-process HTTPS fixture to force HTTP/1.1.

Verification

The updated branch was built and tested locally with .NET 10.0.101. Existing compiler, analyzer, and obsolete-API warnings remain; no new errors were introduced.

• Debug focused parser/framing/pool/upgrade/keep-alive set: 66 passed, 0 failed, 0 skipped.
• Release focused parser/framing/pool/upgrade/keep-alive set: 66 passed, 0 failed, 0 skipped.
• Debug complete Fluxzy.Tests.UnitTests.Core set: 189 passed, 0 failed, 0 skipped.
• Release complete Fluxzy.Tests.UnitTests.Core set: 189 passed, 0 failed, 0 skipped.
• Debug complete Fluxzy.Tests.UnitTests.H11Client set: 10 passed, 0 failed, 0 skipped.
• Independent final read-only review found no remaining medium-or-higher correctness issues after the transfer-framing and status-205 follow-up.
• The real TLS test verifies that two HTTP/1.1 requests 1.25 seconds apart reuse one upstream connection.
• Git whitespace validation passes with the repository's CRLF convention (core.whitespace=cr-at-eol).

The complete repository test suite was not run because it includes environment-dependent external-network, certificate-installation, and raw-capture cases. The test startup reported the expected inability to install its development CA in the unprivileged DevPod; the selected tests still completed successfully.

Line Count