feat: adopt Go 1.27 (encoding/json/v2, jsontext, generic methods)

[mpyw]

Raises the toolchain to Go 1.27 and takes up the parts of 1.25–1.27 that this project can actually use.

encoding/json/v2 — request bodies

internal/server/body/json.go decodes with v2 defaults, which follow RFC 7493. The parsed map becomes the SQL bind parameters, so v1 silently keeping the last of a duplicate member meant {"id":1,"id":2} could override a parameter a caller believed it had set. Duplicate members and invalid UTF-8 are now rejected as 400.

encoding/json/v2 — responses

Responses are this project's public API, so marshalOptions opts out of v2's stricter/tidier defaults one by one rather than accepting them wholesale. internal/server/responder_test.go pins every row below, so a failure means a v2 default leaked into the wire format:

	v1	v2 default	this PR
map key order	sorted	insertion	sorted (Deterministic)
non-UTF-8 column value	U+FFFD	error → 500	U+FFFD (AllowInvalidUTF8)
nil slice	null	[]	null (FormatNilSliceAsNull)
nil map	null	{}	null (FormatNilMapAsNull)
time.Time	RFC 3339	RFC 3339	RFC 3339
[]byte column	base64	base64	base64
NaN / Inf	error	error	error → 500
< > &	< …	literal	literal

With those four options, v2 output is byte-identical to v1 for every type a SQL driver can hand back. The last row is the only remaining difference and it is deliberate: responses are served as application/json, where v1's HTML escaping bought nothing.

Two of these would have been silent regressions: without AllowInvalidUTF8, rows that serve fine today (latin1/binary columns) become a 500; the nil-map case is reachable through a post-transform handing back a nil map for an empty type: one result.

Mock JSON/JSONL sources decode with v2 too, so a duplicate key in fixture data is now a config error instead of a silent last-wins.

encoding/json/jsontext — JSONL streaming

JSONL parsing streams through a jsontext.Decoder instead of reading whole lines with bufio.Scanner. This removes the 64KiB per-line limit that made any large JSONL row fail with token too long; a regression test covers a 128KiB row.

Generic methods

internal/mock/filter.go and internal/mock/value_parser.go each carried their own copy of the pooled-goja-VM dance — sync.Pool wiring, VM construction, timeout watchdog, interrupt unwrapping. Both now use js.PooledVM, whose Call method takes the result type as a parameter:

p.Call(map[string]any{"ctx": ctx}, goja.Value.ToBoolean, row, input)  // bool
p.Call(nil, goja.Value.Export, value)                                 // any

The type parameter belongs on the method rather than a free function because the conversion has to run while the runtime is still checked out — a goja.Value must not outlive its runtime's return to the pool.

Bug found while consolidating

goja.Runtime.Interrupt leaves its flag set if it lands after the script already finished, and time.Timer.Stop does not wait for an in-flight timer function. A filter row finishing right at the 5s deadline could therefore return a poisoned runtime to the pool, and every later row on that runtime failed with a spurious timeout.

runWithTimeout now disarms the watchdog under a mutex and clears the flag before returning, and a runtime that really did time out is dropped instead of pooled. TestRunWithTimeout_ClearsLateInterrupt fails without the fix (verified by reverting it).

Other new API applied

API	Since	Where
errors.AsType[E]	1.26	all 5 errors.As sites — 4 used the (*T)(nil) idiom just to get a typed target
http.Server.MaxHeaderValueCount	1.27	100, with MaxHeaderBytes 64KiB and a ReadHeaderTimeout
url.URL.Clone	1.27	js.NewRequest, so the request JS sees is a snapshot
strings.CutLast	1.27	config.getParentPath
testing/synctest.Test	1.25	the delay tests
maps.Copy	—	CSV row copy (via go fix)

Only the header phase is bounded by a timeout, so large uploads and slow queries are not cut off.

Checked and not applied

The remaining new API in 1.25–1.27 was cross-referenced against the official api/go1.{25,26,27}.txt lists, filtered to the packages this repo imports directly. Nothing else fits — sync.WaitGroup.Go (no WaitGroup), database/sql.ConvertAssign (sqlx MapScan does the conversion), httptest.NewTestServer (tests call handlers directly), bytes.CutLast, bytes.Buffer.Peek, slog.NewMultiHandler, os.Root methods.

One deliberate omission: net/http.NewCrossOriginProtection (1.25) would add Sec-Fetch-Site-based CSRF protection, which is genuinely relevant for a proxy whose POST/PUT/PATCH/DELETE endpoints run SQL. But it is a policy decision that interacts with the configurable permissive CORS, not a mechanical toolchain update — worth its own PR if you want it.

Tests

• decodeJSON[T] replaces 79 hand-rolled decode blocks across the e2e suite. Every assertion routed through it now also asserts the response is well-formed RFC 7493 JSON.
• The delay tests run under testing/synctest: they assert the exact configured delay instead of a tolerance window, and no longer sleep (0.00s per subtest, and the flaky < 30ms assertion is gone).
• New: internal/server/responder_test.go (the compatibility table above, plus the NaN 500 path and that Error's lo.Must cannot panic on a non-UTF-8 driver message), the strict decode paths, the 128KiB JSONL row, and the interrupt fix.

Verified

• go build ./..., go vet ./..., gofmt -l (Go 1.27.0) — clean
• go test ./... and go test -race ./... — all pass
• golangci-lint 2.13.1 (built with go1.27.0) — 0 issues

Notes for review

• golangci-lint must be ≥ v2.13.0 to parse generic methods (go1.27 support landed in v2.13.0, 2026-08-19). golangci-lint-action@v9 resolves to the latest release, so CI is fine unpinned — say the word if you'd rather pin it.
• CI's test matrix drops to ['1.27']; encoding/json/v2 and generic methods rule out older toolchains. README/CLAUDE.md updated.
• GOEXPERIMENT=nojsonv2 will no longer build this project. That opt-out is temporary upstream.
• gofmt from a Go ≤1.26 toolchain cannot parse internal/js/pool.go (generic method). Contributors need a 1.27 toolchain locally.

🤖 Generated with Claude Code

[codecov]

Codecov Report

❌ Patch coverage is 90.90909% with 10 lines in your changes missing coverage. Please review.
✅ Project coverage is 82.90%. Comparing base (7bc1246) to head (74ac70b).

Files with missing lines	Patch %	Lines
internal/cli/commands/app.go	0.00%	5 Missing ⚠️
internal/js/pool.go	93.54%	4 Missing ⚠️
internal/mock/json.go	92.30%	1 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##             main      #18      +/-   ##
==========================================
+ Coverage   76.55%   82.90%   +6.35%     
==========================================
  Files          34       35       +1     
  Lines        2013     1685     -328     
==========================================
- Hits         1541     1397     -144     
+ Misses        346      288      -58     
+ Partials      126        0     -126     

Files with missing lines	Coverage Δ
internal/config/config.go	86.66% <100.00%> (+3.57%)	⬆️
internal/config/errors.go	94.68% <100.00%> (+0.98%)	⬆️
internal/js/apply.go	90.36% <100.00%> (+5.04%)	⬆️
internal/js/error.go	88.46% <100.00%> (+3.16%)	⬆️
internal/js/request.go	100.00% <100.00%> (ø)
internal/mock/csv.go	93.33% <100.00%> (+5.83%)	⬆️
internal/mock/filter.go	77.14% <100.00%> (+26.54%)	⬆️
internal/mock/value_parser.go	75.00% <100.00%> (+29.34%)	⬆️
internal/server/body/json.go	100.00% <ø> (ø)
internal/server/handler.go	89.53% <100.00%> (+0.31%)	⬆️
... and 4 more

... and 21 files with indirect coverage changes

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.