ref: remove experimental_prefetchInRender - #11221

Replaces the ten open weekly Dependabot PRs with one reviewed, tested
set. All ten were included; nothing was held or excluded.

## Screening

- **Screening time:** `2026-09-09T06:44:31Z`
- **72-hour cutoff:** a version must have been published before
`2026-09-06T06:44:31Z`

Soak is measured from **upstream publication time** (npm registry `time`
metadata; `proxy.golang.org` `.info` metadata), never from Dependabot PR
creation — all ten PRs were opened ~2h before screening and would fail a
naive PR-age check. The gate was re-applied to the **final resolved
graph**, including every transitive addition and upgrade.

## Included

| PR | Dependency | Old → New | Published (UTC) | Soak | Risk conclusion
|
|---|---|---|---|---|---|
| #1697 | `github.com/klauspost/compress` | 1.19.2 → 1.20.0 |
2026-09-02T12:08:18Z | 6d 18h | Radar imports only `compress/gzip`
([`internal/server/compress.go`](internal/server/compress.go)). Release
is additive (lzw, XPRESS, arm64 asm) plus
[#1201](klauspost/compress#1201), a
**correctness fix** backporting
[golang/go#80538](golang/go#80538) for flate
L7–9 dictionary streams. Radar uses no dictionaries; `gzhttp` jitter
change is in a package Radar doesn't import. |
| #1698 | `modernc.org/sqlite` | 1.57.0 → 1.58.0 | 2026-09-01T15:17:17Z
| 7d 15h | [SQLite 3.53.4](https://sqlite.org/releaselog/3_53_4.html);
upstream's own journal-rollback corruption fix replaces the local
super-journal patch, recovery behavior unchanged. New OFD locking is
**opt-in and off by default** — Radar sets neither
`MODERNC_SQLITE_OFD_LOCK` nor `OFDLocking()`, so locking is
byte-for-byte unchanged. `isCorruptedSQLiteError` in
[`internal/timeline/manager.go`](internal/timeline/manager.go) matches
SQLite core result text; its tests pass. |
| #1699 | `golang.org/x/sys` | 0.47.0 → 0.48.0 | 2026-08-31T19:43:43Z |
8d 11h | Syscall constants/types and cpu detection only. Radar's use is
`windows` (localterm, process) and `unix` (birthtime). Verified by
cross-compiling windows/linux/darwin. |
| #1700 | `github.com/coreos/go-oidc/v3` | 3.20.0 → 3.21.0 |
2026-09-01T23:03:42Z | 7d 7h | Single change:
[#499](coreos/go-oidc#499) ignores JWKs with
unsupported key types instead of failing the whole key set. Tolerance
improvement, not a verification weakening — a supported key is still
required to verify signatures. Auth surface is
[`internal/auth/oidc.go`](internal/auth/oidc.go); `internal/auth` tests
pass. |
| #1701 | `github.com/prometheus/common` | 0.70.1 → 0.71.0 |
2026-08-31T08:47:38Z | 8d 21h | Release is mostly OpenMetrics 2.0
encoder/negotiation work. Radar touches neither: it calls only
`expfmt.NewTextParser(model.LegacyValidation)`
([`internal/upgrade/collectors_live.go`](internal/upgrade/collectors_live.go))
and `model.ParseDuration`
([`internal/mcp/tools_prometheus.go`](internal/mcp/tools_prometheus.go)).
`model` changes are additive/perf. |
| #1702 | `postcss` | 8.5.26 → 8.5.28 | 2026-09-03T15:13:59Z | 5d 15h |
8.5.27 fixes comment handling for `/*#`, `*` hack in custom properties,
and `list.comma()`/`list.space()` edge cases; 8.5.28 fixes the types
regression 8.5.27 introduced (so landing on .28, not .27, matters). CSS
output verified via `make build` + visual test. |
| #1703 | `@tanstack/react-query` | 5.101.4 → 5.102.8 |
2026-08-27T16:06:57Z | 12d 14h | Widest range here. 5.102.0 is a minor
with real **removals** — the `promise` property on query results and
render-time prefetching
([#11221](TanStack/query#11221)),
`experimental_beforeQuery`/`experimental_afterQuery`
([#11233](TanStack/query#11233)), and
`placeholderData` on suspense infinite queries
([#11144](TanStack/query#11144)). Grepped Radar
for each: **zero call sites** for all of them, and no suspense/infinite
query hooks at all. `placeholderData` is used only with plain
`useQuery`. Type changes covered by a clean `make tsc`. |
| #1704 | `react-router-dom` | 7.18.2 → 7.18.3 | 2026-08-28T14:30:10Z |
11d 16h |
[v7.18.3](https://github.com/remix-run/react-router/blob/v7/CHANGELOG.md#v7183):
route-matching perf plus two hardening changes.
[#15446](remix-run/react-router#15446) makes
programmatic navigations **reject external destinations** — checked all
70 `navigate()` call sites in `web/src` and `packages/k8s-ui/src`; every
one passes a relative path or a `{ pathname, search }` object, none an
external URL. |
| #1705 | `react-virtuoso` | 4.18.12 → 4.18.13 | 2026-09-05T14:29:31Z |
**3d 16h** | Closest to the cutoff, still clears it. Single patch
([#1493](petyosi/react-virtuoso#1493)) fixing a
one-frame blank when an older page is prepended — directly in the
log-viewer path. Log viewer captured in the visual test. |
| #1706 | `eslint` | 10.8.0 → 10.10.0 | 2026-09-04T14:34:21Z | 4d 16h |
Dev-only, but the **largest graph churn** — see below. |

## Graph movement beyond the named targets

Every item below was independently age-checked against the same cutoff
and attributed to a named cause.

**Go — lockstep and upstream requirement bumps:**

| Module | Old → New | Published | Why |
|---|---|---|---|
| `modernc.org/libc` | 1.74.4 → 1.75.6 | 2026-08-26 | **Required
lockstep.** `modernc.org/sqlite` v1.58.0 explicitly requires downstreams
to pin the exact `libc` it pins ([issue
#177](https://gitlab.com/cznic/sqlite/-/issues/177)). Verified: sqlite's
`go.mod` pins `libc v1.75.6` / `memory v1.12.1`, and ours resolves to
exactly those. |
| `modernc.org/memory` | 1.11.0 → 1.12.1 | 2026-08-19 | Same lockstep. |
| `modernc.org/cc/v4` | 4.29.1 → 4.29.2 | 2026-07-28 | Build-time
transpiler deps of libc. |
| `modernc.org/ccgo/v4` | 4.34.6 → 4.35.0 | 2026-08-06 | Same. |
| `modernc.org/gc/v3` | 3.1.4 → 3.1.5 | 2026-06-24 | Same. |
| `github.com/stretchr/testify` | 1.11.1 → 1.12.1 | 2026-08-17 | Pulled
by `prometheus/common`. Test-only. Fixes plus a swap to
`go.yaml.in/yaml/v3`. |
| `go.yaml.in/yaml/v3` | 3.0.4 → 3.0.5 | 2026-07-26 | Pulled by
`prometheus/common`/testify. Reviewed the full commit range:
test-infrastructure cleanup, a gofmt revert, and retraction of
uninstallable v3 tags. **No parser behavior change.** |

**npm — all from eslint's `file-entry-cache` v11 update
([eslint#20801](eslint/eslint#20801

`file-entry-cache` 8.0.0 → 11.1.5, `flat-cache` 4.0.1 → 6.1.23, `keyv`
4.5.4 → 5.6.0 (three major jumps), `flatted` 3.4.2 → 3.4.4,
`@eslint/plugin-kit` 0.7.2 → 0.7.3; **added** `@cacheable/memory`,
`@cacheable/utils`, `@keyv/bigmap`, `@keyv/serialize`, `cacheable`,
`hashery`, `hookified` (1.15.1 + 2.2.0), `qified`; **removed**
`json-buffer`. Oldest addition is `@keyv/serialize` (2025-09-10), newest
`nanoid` 3.3.18 (2026-08-07, via postcss) — all far past the cutoff.
This is eslint's `--cache` layer only; it is not in any runtime path and
not in the shipped bundle.

This churn is the one thing in the batch I'd flag for a second look. It
is well-soaked, fully attributed to a single deliberate upstream change,
and dev-only — but it is three major transitive jumps to buy two eslint
patch releases. Happy to drop #1706 and keep the other nine if you'd
rather not carry it this week.

## Verification

Run against the batch as a whole, from a clean worktree on
`origin/main`:

- `make tsc` — **clean.** Load-bearing here: react-query 5.102.0 changed
several public types (`NoInfer` revert
[#11245](TanStack/query#11245), `queryOptions`
return types [#11224](TanStack/query#11224),
`UseInfiniteQueryOptions` TData default
[#11147](TanStack/query#11147)) and postcss
8.5.28 fixes a types regression.
- `make test` — **all packages pass, 0 failures.** Consumer packages
re-run uncached (`-count=1`): `internal/auth` (go-oidc),
`internal/timeline` (sqlite, incl. the corruption-marker tests),
`internal/upgrade` (expfmt), `internal/ai` (sqlite), `internal/mcp` —
all ok.
- `make build` — **succeeds** end to end (frontend → embed → binary).
- `go mod verify` — **all modules verified.**
- Cross-compile — `windows/amd64`, `windows/arm64`, `linux/amd64`,
`linux/arm64`, `darwin/arm64` all build (targeted at the `x/sys` bump).
- **visual-test: ran** · 3 shots ·
`/Users/nadaverell/workspace/sky/ws5/radar/.playwright-mcp/visual-test/depsbatch-20260909/`
— run against `kind-radar-gpu-ecosystem-demo`. Not skipped, because
postcss (CSS output), react-virtuoso (list rendering), react-router
(routing) and react-query (data layer) can all materially change what
renders. Overview, Resources/Pods table, and the pod **log viewer** (the
virtuoso prepend path) all render correctly; console clean apart from an
environmental `404 /api/metrics/...` (demo cluster has no
metrics-server).

Two pre-existing conditions confirmed identical on `origin/main` and
unrelated to this batch: 4 `go vet` "copies lock" warnings in
`internal/server` test files (CI doesn't run vet), and a `darwin/amd64`
build failure in `cmd/desktop` (cgo/build-tag issue with
`startNativeMouseMonitor`).

`eslint` is the one dependency whose targeted test I did **not** run
locally — `npm run lint` is flagged broken in `.claude/commands/qa.md`.
CI does run it (`.github/workflows/ci.yml` "Lint" step), so the eslint
bump is verified there instead, as a **before/after comparison**:

| eslint | Commit | Result |
|---|---|---|
| 10.8.0 | `750cea6c` (this PR's base,
[run](https://github.com/skyhook-io/radar/actions/runs/34225225175)) | ✖
353 problems (**0 errors**, 353 warnings) |
| 10.10.0 | this PR
([run](https://github.com/skyhook-io/radar/actions/runs/34322370732/job/102371700492))
| ✖ 353 problems (**0 errors**, 353 warnings) |

Identical. 10.9.0/10.10.0 fix several rule false negatives (`new-cap`,
`no-extra-bind`, `no-loss-of-precision`, `no-unexpected-multiline`), so
new findings were possible — none materialised on Radar's source. No
source change was needed to land the bump.

## CI

All checks green on the batch: Backend, Frontend, pkg (shared library),
k8s-ui (shared UI), Helm chart, CodeQL (go + javascript-typescript),
Cursor Bugbot (no findings).

## Scope

Dependency manifests only — `go.mod`, `go.sum`, `package-lock.json`,
`web/package.json`, `packages/k8s-ui/package.json`. No application
source, generated frontend bundle, workflow, tag, or publishing config
changed. `git diff --check` clean; the root `dompurify` override is
untouched. No PR was stale/already-landed, and nothing required a Radar
source change or compatibility shim.

Supersedes #1697, #1698, #1699, #1700, #1701, #1702, #1703, #1704,
#1705, #1706.

🤖 Generated with [Claude Code](https://claude.com/claude-code)

https://claude.ai/code/session_01XArawWJfTECJQAjnWG9L1X

Co-authored-by: Claude Opus 5 (1M context) <[email protected]>