fix(build): unbreak sketchcollector OCB build against current sketchlib-go

[zzylol]

Summary

Two root causes when running `build_sketchcollector.sh` against today's sketchlib-go — one tooling-glue (the actual blocker), one real API drift in two processors.

(1) sketchlib-go was resolving via the proxy, not the local checkout

The OCB step writes `cmd/sketchcollector/go.mod` with `sketchlib-go ... // indirect` in the require block. The build script's "if no replace, append one" check was `grep -q "sketchlib-go"`, which matches the require line, short-circuits the `if`, and never appends the replace. So `go build` resolved sketchlib-go via `sum.golang.org`, pinning to a months-stale published commit (`20260328`) that lacked the methods the processors call.

That's why `InsertValue`, `EstimateCardinality`, `SerializeMsgpack` etc. came back "undefined" even though they exist on `sketchlib-go@main`.

Fix in two places:

• `builder-config.yaml` gets a `sketchlib-go` entry in `replaces:` so OCB writes the replace into the generated `go.mod` from the start. Path is `../../../../sketchlib-go` relative to `cmd/sketchcollector/go.mod`, which lands on the sibling repo.
• `build_sketchcollector.sh` check now matches the replace line specifically (`^replace[[:space:]]+...`) — defensive backup.

(2) Real API drift — two processors

After (1), the build still fails on:

`ddsketchprocessor`: `DeserializeState` / `DeserializePortable` are gone

sketchlib-go #52 (`cbbe94c`) renamed the inbound constructors:

• `DeserializeState(&state)` → `NewFromState(&state)`
• `DeserializePortable(&env)` folded into `NewFromState(env.GetDdsketch())` (extract the oneof variant)

The bare-state fallback swaps the function name. The envelope path uses the generated `GetDdsketch()` accessor (returns nil safely if upstream sent a different variant — guarded with a clear error).

CMS / CountSketch merge processors: `ApplyDelta` signature changed

sketchlib-go's `ApplyDelta` for both CMS and CountSketch is now `(*X, *Delta)` (no return), not `(*X, []byte) error`. Callers must call `DeserializeDelta(payload)` first. Updated both merge processors to do that and surface decode failures with the same logger/zap pattern they used before.

(`DDSketch.ApplyDelta` still takes `[]byte` and returns error, but `ddsketchprocessor` doesn't actually call it today — its delta-encoding inbound path is stubbed pending the go-side delta encoder. No change there.)

Verification

`./build_sketchcollector.sh --skip-patches` now produces a clean binary:

```
$ ls -la opentelemetry-collector-contrib-patch/cmd/sketchcollector/sketchcollector
-rwxrwxr-x ... 494990016 ... sketchcollector

$ ./sketchcollector components | grep -E 'name: (HLL|KLL|ddsketch|countmin|countsketch|countminsketchmerge|countsketchmerge)'
- name: HLL
- name: KLL
- name: countmin
- name: countminsketchmerge
- name: countsketch
- name: countsketchmerge
- name: ddsketch
```

All seven sketch processors compiled in.

`Dockerfile.sketchcol`'s "build prereqs" comment block was tracking the just-in-time replace approach; updated to point at the sibling `../sketchlib-go` checkout (no manifest edit needed at build time anymore).

Test plan

• `build_sketchcollector.sh --skip-patches` exits 0 against current `sketchlib-go@main`.
• `sketchcollector components` lists all 7 sketch processors.
• `docker build -f deploy/docker/Dockerfile.sketchcol -t asap/sketchcol:dev .` and re-run e2e — should now exercise the real merge processors and the patched ddsketch decode path. (The agent-side ddsketch→batch wiring issue I flagged in fix(gateway): unbreak boot post-#206 + simplify merge config #207's PR description is independent of this.)

🤖 Generated with Claude Code