feat: adopt Go 1.27 JSON v2 semantics

[mariusvniekerk]

Moves AgentsView to Go 1.27 and adopts native encoding/json/v2 behavior
throughout. This is a clean migration with no JSON v1 compatibility or
checkpoint fallback paths.

Typed Huma routes now share JSON v2 request and response semantics with the
rest of the application. OpenAPI continues to describe token usage as
arbitrary JSON. JSONL rewrite paths retain untouched values as raw JSON, so
large integers stay exact, while stored map output is deterministic where bytes
define artifact identity.

The Go 1.27 audit also adopts generic methods, assignment-context inference,
promoted fields in struct literals, errors.AsType, and related
standard-library simplifications. Build images, local tooling, documentation,
and reviewer guidance now require Go 1.27. golangci-lint is pinned to v2.13.0,
which can analyze the new language level.

[roborev-ci]

roborev: Combined Review (a0389b2)

High-severity compilation failure and medium-severity nondeterministic persisted JSON need fixing before merge.

High

• internal/db/messages.go:112 — Changing Message.TokenUsage to jsontext.Value breaks test fixtures that still assign the distinct type encoding/json.RawMessage, including internal/db/usage_cache_fill_test.go:29 and internal/postgres/usage_facts_parity_pgtest_test.go:188. The test packages will not compile. Convert all remaining assignments and imports to jsontext.Value, or retain a source-compatible field type.

Medium

• internal/parser/workbuddy.go:302 — JSON v2 does not sort map keys by default, so persisting map-backed JSON in TokenUsage or InputJSON can produce different byte sequences when identical input is reparsed. Byte-for-byte parsed-diff and fingerprint comparisons may then report false changes and trigger unnecessary synchronization. Similar sites include gptme.go:204, vscode_copilot.go:593, visualstudio_copilot.go:1316, and roocode.go:1335. Use json.Deterministic(true) or otherwise canonicalize map-backed JSON before persistence.

---

Reviewers: 2 done | Synthesis: codex, 14s | Total: 12m45s

[roborev-ci]

roborev: Combined Review (f3044da)

High-severity persisted-hash compatibility regression and three medium-severity determinism, compatibility, and memory issues remain.

High

• Persisted hashes change without a version bump — internal/db/recall_evidence_window.go:507, internal/artifact/canonical_json.go:12, internal/rawsync/manifest.go:185

  Migrating from the previous JSON representation to JSON v2/JCS changes hashes for strings containing characters such as <, >, or &, while format and digest versions remain unchanged. This can revoke provenance_ok for unchanged evidence, cause raw-ingest retry conflicts, and assign new artifact identities to unchanged content.

  Fix: Preserve legacy encoding for existing versions, or introduce new versions with dual-version validation and migration. Add compatibility vectors containing HTML-sensitive characters.

Medium

• Zero numeric fields now alter pinned segment bytes — internal/artifact/wire.go:218, internal/artifact/wire.go:231

  These fields still use omitempty, but JSON v2 no longer omits zero numbers. Empty tool-call or result events can therefore change content hashes without a segment-version bump.

  Fix: Change both tags to omitzero and add a compatibility test for zero-length nested tool results.

• Persisted parser JSON is nondeterministic — internal/parser/kilo_legacy.go:1755, internal/parser/kilo_legacy.go:1767, internal/parser/poolside.go:358, internal/parser/zed_helpers.go:163

  Maps are serialized into persisted InputJSON without deterministic marshaling. Reprocessing identical data may produce different bytes and ToolCallFingerprint values, causing spurious session changes and mirror pushes.

  Fix: Use deterministic marshaling for all parser-generated or normalized JSON, and test repeated parsing for identical input and fingerprints.

• Checkpoint bootstrap can consume excessive memory — internal/artifact/export_checkpoint.go:173

  The implementation reads and clones up to 64 MiB, unmarshals the complete sessions map, and serializes it again, potentially requiring several hundred megabytes and risking OOM.

  Fix: Restore streaming validation and hashing with jsontext.Decoder, retaining one session entry at a time while preserving the decoded-size limit.

---

Reviewers: 2 done | Synthesis: codex, 19s | Total: 20m51s

[roborev-ci]

roborev: Combined Review (27d14b6)

Migration is generally sound, but two medium-severity issues remain.

Medium

• Non-deterministic persisted tool inputs — internal/parser/kilo_legacy.go:1755, internal/parser/poolside.go:358
  Map-backed tool inputs are marshaled without json.Deterministic(true). JSON v2 does not sort map keys by default, so reparsing unchanged sessions can produce different InputJSON, causing spurious revisions and inconsistent exports. Enable deterministic marshaling and test repeated and full-versus-incremental parsing stability.

• Excessive checkpoint bootstrap memory use — internal/artifact/export_checkpoint.go:173
  Bootstrap now reads, clones, canonicalizes, fully unmarshals, and re-marshals checkpoints up to 64 MiB. Multiple full-size buffers and the decoded session map can coexist, regressing streaming behavior and risking excessive memory use or OOM for valid large archives. Restore streaming validation and incremental hashing to avoid full-body copies and map materialization.

---

Reviewers: 2 done | Synthesis: codex, 13s | Total: 15m29s