fix: parse multi-unit duration strings in convert_duration_to_interval

[juanuicich]

Summary

convert_duration_to_interval(duration text) only inspects the leading integer and the final unit suffix. Multi-unit Go duration strings such as 42m30s therefore parse as roughly 42s. Registration accepts the value through time.ParseDuration and the engine stores the original string, so timeouts silently shrink at enforcement time.

This PR replaces the helper with one that walks the whole string and sums every component into a single fixed interval.

Fixes #3894

What changed

• New migration 20260512135156_v1_0_106.sql does CREATE OR REPLACE FUNCTION convert_duration_to_interval(duration text).
• The replacement parser supports the units the previous helper accepted (ms, s, m, h), decimal components, and an optional leading sign.
• Legacy single-unit suffixes d, w, y keep their prior calendar semantics. They are only accepted as the entire input ('1d', '1y'), so mixed forms such as '30s1d' fall through to the standard fallback.
• Invalid input returns the existing fallback interval, 5 minutes, matching the live function in 20240321215205_v0_17_1.sql.
• The -- +goose Down block restores the body of that prior migration verbatim.
• frontend/docs/pages/v1/timeouts.mdx and frontend/docs/pages/v1/error-handling/timeouts.mdx describe the multi-unit, decimal format that the engine now enforces.

Behavior changes

• Decimal single-unit values are now parsed correctly. The previous helper truncated 1.5h to 1h because num_value was INT. The new helper returns 1h30m, which matches what registration validation already accepted.

The fix is engine-side because stored values and other clients can already send multi-unit strings. An SDK-only change would not cover them.

Regression coverage

pkg/repository/duration_test.go adds a database-facing test that runs the migration through setupPostgresWithMigration(t).

Multi-unit and fractional cases compare seconds via:

SELECT EXTRACT(EPOCH FROM convert_duration_to_interval($1))::double precision

Cases:

• 42s to 42
• 11m to 660
• 1h to 3600
• 42m30s to 2550
• 1h30m to 5400
• 1h30m5s to 5405
• 1500ms to 1.5
• 1.5h to 5400
• -1.5h to -5400
• 0s to 0
• bad, 42, 30s1d to 300 (the 5 minutes fallback)

Legacy single-unit cases compare interval values directly to confirm calendar semantics are preserved:

• 1d equals 1 day
• 1w equals 7 days
• 1y equals 12 mons
• 10d equals 10 days

Generated files

task generate-sqlc produces no diff. convert_duration_to_interval is not declared in the sqlc schema files, so no query signatures change.

Test plan

• go test -count=1 ./pkg/repository -run TestConvertDurationToInterval -v: passing
• go test -count=1 ./pkg/validator -run TestValidatorDuration -v: passing
• task generate-sqlc produces no diff
• gofmt -s -l on changed files reports no diff (the task fmt-go body)
• task test: full unit suite, 33 packages with tests, all pass
• task lint-go: golangci-lint run ./... --config .golangci.yml reports 170 pre-existing issues, none in files this PR touches
• task test-integration: go test -count=1 -tags integration $(go list ./... | grep -v quickstart) -failfast against docker compose Postgres + RabbitMQ, all pass

[vercel]

@juanuicich is attempting to deploy a commit to the Hatchet Team on Vercel.

A member of the Team first needs to authorize it.