Optimize protobuf encoding and add CI performance reports - #4

Protobuf WKT encoding performs avoidable reflective reads, and numeric map keys allocate temporary strings. This PR uses concrete WKT getters and direct quoted numeric-key writes, fixes codec edge cases, and adds repeatable CI comparisons so future PRs and default-branch pushes report their performance impact.

Encoding changes

• Use concrete Timestamp/Duration/Struct/Value/ListValue getters with DynamicMessage fallback, one exact-size uint64 buffer, cached descriptors/WKT classification, and presence checks before reflective reads.
• Fix deprecated fields omitted by typed encoding, unsigned map keys printed as signed values, and custom JSON names producing invalid JSON or uncompilable generated codecs.
• Preserve numeric/boolean map-key spelling under writer settings; add regression fixtures and a 30-case codegen/typed/reflection × UTF-16/UTF-8 JMH matrix.

The full same-compiler CI comparison on Java 21/25 measured Struct throughput +95–116%, codegen/typed maps +12–29%, and reflection Timestamp +25–29%. Struct allocations fell about 44%; map allocations fell 26–39% across all paths. Of 60 comparisons, 31 show improvement signals, 29 are inconclusive, and none show regression signals or allocation growth alerts. These are per-workload advisory signals, not an overall application speedup. Full CI results and raw artifacts supersede the earlier mixed-compiler follow-up; the historical local reports and compact CSV evidence remain in docs/performance-port-2026-09-05.md and docs/performance-results/.

Performance CI

• Compare exact PR base/head SHAs and pre-/post-push SHAs on main, using stock JMH on Java 21 and 25.
• Clean-build both revisions with the same compiler, shared candidate benchmark inputs, and isolated Maven repositories; measure sequentially on the same runner, alternating variant order.
• Retain raw samples, environment, source identities, build logs and Markdown reports for 90 days. Job summaries are available on every run.
• A separate trusted workflow_run publisher validates JSON artifacts, adds commit checks, and updates one PR comment while rejecting stale results. It never executes PR/artifact code with write credentials.
• Throughput alerts require ≥10% change and separated JMH intervals and remain advisory. The 14 allocation budgets stay enforced; missing/nonfinite GC measurements now fail and allocation reports are uploaded.

The publisher starts automatically once its workflow exists on main. This introducing PR already gets measurement job summaries and artifacts. No external account, JMH fork, branch-protection change, or paid runner is required. docs/performance-ci.md compares GitHub-native reporting, dedicated runners, CodSpeed, Bencher, and github-action-benchmark and explains protocol/limits.

Validation

• Encoding: clean Maven verification on Corretto 21.0.11 and 25.0.4; 496 tests passed on each with -Werror.
• Current commit passed build, all 14 allocation checks, and the official protobuf conformance suite: CI run.
• CI tooling: 15 tests cover measurement validation, thresholds, invalid allocation data, artifact bounds/identity, token redirects, and stale/missing PR reports.
• End-to-end local smoke run clean-built main and candidate and successfully measured three cases with identical JDK/input sources; smoke output is explicitly excluded from performance evidence.
• Actionlint, shell syntax checks, Spotless, and whitespace checks passed.
• Full Java 21/25 CI comparisons passed on this revision: 30 cases per JVM, all 60 summaries independently reconstructed from raw JMH result files and validated. Performance run.

Context reuse, packed field names, primitive-list rewrites, and CodSpeed infrastructure remain deferred.