fix: report buffered progress when message writes fail - #1033

What type of PR is this?

• Bug Fix

Description

messageWriter.Write and WriteString currently return (0, err) if a later buffer flush fails, even when the current call has already copied a prefix into the message buffer and an earlier frame has reached the peer. This also makes io.Copy report zero progress for that call.

Return nn - len(p) on the buffered-loop error path, so the count reflects bytes accepted from the current argument. Bytes pending from an earlier call are not included. The returned count describes buffer acceptance; it does not guarantee that every counted byte reached the peer. Error identity, sticky failure behavior, successful writes, and the existing large-server fast path are unchanged.

Add a regression matrix using actual HTTP Upgrade/Dial and TCP client/server sockets, with a deterministic write-error wrapper after the first successful write. It covers Write, WriteString, io.Copy, primed-buffer/current-call accounting, separate peer-delivery assertions, zero progress, sticky errors, success, and the large-server fast path. Every socket has bounded deadlines and cleanup.

Related Tickets & Documents

The original-source run reproduces the count error: original RED, unchanged production blob 9562ffd4978ccabea0da2f06256e1c281bcefb7f. Five existing controls passed; twelve new progress cases failed (0 rather than 16/13), while all fourteen boundary/control cases passed. The same regression source is used in this PR.

Existing proxy-test instability is tracked separately in #982; no proxy-test change is included here.

Added/updated tests?

• Yes

Run verifications and test

An isolated public-fork Ubuntu workflow explicitly checked out exact source commit 97f927d22d5982ce81101efe76adac24bf59fd78 (native validation). The workflow commit differs from the source commit and is excluded from this PR.

• Go 1.20.14, 1.21.13, 1.22.12 and 1.27.1: all 26 new socket regressions, existing focused controls, complete go test -v -race -coverprofile=... ./..., go vet ./... and formatting of both changed files passed. The full run has 77 passing top-level tests and 30 passing subtests; no race warnings or skips.

• Go 1.22: all 20 existing benchmark cases executed with -benchtime=1x -benchmem. This is a smoke check, not a performance claim.

• Both changed-file Git blobs/SHA256 and actual source HEAD were verified before and after all commands, with zero tracked diff. Both writer methods have 100% statement coverage in this run.

• Full original/fixed race suites were paired. Earlier attempts encountered existing proxy-test flakes, and the final Go 1.27 baseline also failed an existing proxy test while the fixed full run passed. These failures are retained; this PR does not claim to repair the proxy flake.

• Whole-project Go 1.27 formatting fails on both original and fixed sources with the same 6127-byte doc.go diff (comparison exit 0). The changed files pass gofmt.

• Org-guide tools were run against original and exact fixed source without suppressions: golint and govulncheck exit 0 on both. Govulncheck reports unreachable dependency advisories, so this is not a claim of no advisories. Gosec exits 1 with the same 55 diagnostics on both. Golangci-lint exits 1 on both; with diagnostic limits removed, all 124 diagnostics are identical, and none concern the new test or modified returns.

• make verify is passing

• make test is passing

This repository has no Makefile or those targets. Both commands were attempted on original and fixed sources and returned exit 2; the equivalent project-native tests and quality commands are reported above. The aggregate helper workflow correctly remains FAILURE because it preserves those existing format/tool failures rather than claiming all checks green.

This change was developed with AI assistance and verified with the project's native commands.

The org-template reviewer-team request for gorilla/pull-request-reviewers was attempted, but GitHub rejected RequestReviewsByLogin because this account lacks permission. No reviewer-team assignment is claimed.