0.2.0: product hardening (versioning, config, typed errors, docs, packaging) - #4

Replace the bare {evaluations, fileErrors} JSON output with a versioned
envelope: outputVersion, trazoVersion, traceSchemaVersion, generatedAt, an
aggregate summary, results, and structured errors. The shape is a documented
contract (docs/output.md) with its own version, so CI and downstream tools can
depend on it and ignore unknown fields across minor bumps.

Introduce report.Summarize as the single source of the aggregate counts, shared
by the JSON summary and the text footer so they cannot disagree. GeneratedAt is
injected via report.Meta, keeping output deterministic in tests.

BREAKING (JSON output): top-level keys evaluations/fileErrors are now
results/errors under the envelope. Update the Python tests that parse the CLI
output and the golden file accordingly.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Add a stdlib-only `config` package and a `-config <path>` flag: a versioned JSON
policy that pins which evaluators run and their thresholds, so a team can commit
one file and get identical evaluation across dev and CI. No new dependency; the
core stays stdlib-only.

- Partial configs are valid (omitted fields keep defaults); unknown fields are
  rejected (a typo is an error, not a silent no-op); version is required.
- Precedence: defaults < config file < explicitly-set flags (via flag.Visit), so
  the config is a reproducible baseline a one-off flag can still override.
- The package is CLI-independent: config.Load + Config.Build return the evaluator
  set for use from another Go program.

Document it in docs/config.md with an example policy, and record it in the
changelog.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

….2.0 M4)

Build the binary once in TestMain and exercise it end to end: clean exit 0, a
bad finding exit 1, a file error exit 2, valid JSON envelope, Markdown FAIL
report, missing path, invalid -format, two-PATH error, version subcommand,
empty directory, and config/flag precedence.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Classify why a file could not be evaluated with runner.ErrorKind (read_file,
invalid_json, invalid_trace, evaluator, canceled) instead of leaving callers to
match on message strings. The kind is set at each failure point in the runner,
emitted as errors[].kind in the JSON output (outputVersion bumped to 1.1, an
additive change), and tagged in the text and validate output.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Add a -verbose flag that prints one metrics line to stderr after a run
(loaded/valid/invalid/evaluated/duration), so batch and CI runs are observable
without parsing the results. A trace-evaluation tool should itself be
observable.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Add a dedicated schema job that runs the Go schema-sync tests and validates the
fixtures and emitter output against the published JSON Schema, so the schema is
an active part of the contract rather than just documentation. Add a Go test
that the shipped example config always loads, keeping the docs from drifting.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Add testdata/complex_run.json: one trace that interleaves two tool calls paired
by id (no orphans), a tool error, a high-cost step, and a node visited enough
times to look like a loop. An integration test runs the full structural
evaluator set over it and asserts how the work divides (tool error -> bad, loop
-> bad, cost overrun -> neutral, terminal end -> no node finding).

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

The M8 integration fixture testdata/complex_run.json is consumed by the Go
integration test; add it to VALID_FIXTURES so the published JSON Schema is
proven to accept it too, closing the cross-language loop for that fixture.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

docs/security.md: what a trace holds (opaque input/output/error payloads), the
one component that sends data off-machine (the LLM judge, final output only),
how to disable network egress, redaction at emit time, and not committing
private traces.

docs/llm-judge.md: the judge's opt-in flags and config keys, what it grades
(last llm_call output, once), determinism (temperature 0, capped tokens), the
timeout/cancellation model, no internal retries, and the deliberate split
between a malformed verdict (neutral finding) and a transport failure
(errors[].kind=evaluator, structural evaluators still run). Notes network-free
testing via the fakeJudge Completion.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

- LICENSE: MIT.
- Makefile: build, check (vet+test), gate (local mirror of the CI evaluator
  gate), install, agent-test, and a cross-compiled `dist` target that produces
  per-platform archives (.tar.gz / .zip) plus a sha256 checksums.txt.
- .github/workflows/release.yml: on a `v*` tag, guard that the tag matches the
  compiled-in const Version, vet+test, `make dist`, then publish the binaries to
  a GitHub release via `gh release create`. No third-party actions; stdlib-only
  ethos extended to CI.
- README: status/release/go-version/license badges, an Install section
  (`go install`, from-source, prebuilt binaries with checksum verification), and
  a Core documentation section linking config/output/versioning/llm-judge/security.
- .gitignore: ignore /bin and a bare ./trazo so ad-hoc builds do not dirty the
  vcs stamp; dist/ is already ignored.
- CHANGELOG: record the integration fixture, the two new docs, and this packaging.

Verified: cross-compiles clean for linux/darwin/windows (amd64/arm64) to static
binaries; `trazo version` and the gate exit codes (0 clean, 1 bad) both correct.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>