fix: remove online serving overview from SDK docs (closes #22)

[9Olive]

Addresses #22: online serving is not yet available for customers, so this removes the section from context/platform/sdk-overview.md (the Online Serving section, the launch_online_serving_endpoint()/predict()/update()/ping()/destroy() examples, and the workflow-step-7 mention) and removes the premature online-serving mentions from context/guides/rfm-vs-training.md's RFM-vs-training comparison table and decision bullets, per the issue's action item to check other customer-facing docs. No other content in either file was touched. When online serving is reintroduced for customers, these sections can be restored following the official docs at https://kumo.ai/docs/fine-tuning/online-serving/.

[9Olive]

Hold: this removes the entire Online Serving section, but the repo authoritatively documents it (the Online Serving section in sdk-overview.md; gap doc-010 endpoint.ping() is marked resolved). Removing it contradicts the repo's own docs. (Aside: kumoai.launch_online_serving_endpoint isn't a top-level symbol in installed 2.22.0, but that's inconclusive - it may live elsewhere.) Please confirm the feature was actually removed upstream before merging.

---

Automated internal-consistency review (cross-checked against the repo's own authoritative docs; a few SDK-surface claims spot-checked against installed kumoai 2.22.0).

[Manushpm8]

Leaving open — this needs a product/maintainer decision, not a source verdict. The Online Serving API this PR removes is fully implemented and shipped in the SDK: launch_online_serving_endpoint(...) (kumo-sdk/kumoai/trainer/job.py:260) returning an OnlineServingEndpoint with predict/ping/update/destroy (kumoai/trainer/online_serving.py). sdk-overview.md is sourced from job.py, so the docs are accurate. The PR's rationale — that online serving isn't yet available to customers — is a product-availability claim that source can neither confirm nor deny. If it's intentionally not GA, gating it with a "not yet generally available" note is preferable to deleting working, accurate API reference; if it is available, the docs should stay. cc @9Olive — your call.

[leena-kang]

Blocked / unable to verify

The SDK proves online serving exists, but cannot establish customer entitlement or GA availability:

# /home/leena/code/kumo/kumo-sdk/kumoai/trainer/job.py:260-264
def launch_online_serving_endpoint(...) -> OnlineServingEndpointFuture:

# /home/leena/code/kumo/kumo-sdk/kumoai/trainer/online_serving.py:37,49,95,100,167
class OnlineServingEndpoint:
    def predict(...):
    def ping(...):
    def update(...):
    def destroy(...):

Accuracy fix — Required to unblock: Obtain confirmation from a maintainer or product owner that online serving is unavailable to customers and should be removed from customer-facing guidance. Repository and SDK evidence alone cannot verify that central claim.

Accuracy fix — Required after that determination: Keep repository metadata consistent. If removal remains, update meta/skills/sync/sync-kumo-sdk.md:33, which still routes online_serving.py into sdk-overview, and context/_gaps.yaml:164-176, where doc-010 remains resolved by documentation that was deleted.

If the documentation is restored, use result(), not attach(), and avoid a numeric latency guarantee:

# /home/leena/code/kumo/kumo-sdk/kumoai/trainer/online_serving.py:171
class OnlineServingEndpointFuture(KumoFuture[OnlineServingEndpoint]):

# /home/leena/code/kumo/kumo-sdk/kumoai/futures.py:42,51,80,85
class KumoFuture:
    def result(...):
class KumoProgressFuture(KumoFuture):
    def attach(...):

Accurate checks

• Removing the deleted attach() example corrects an API error.
• Removing the numeric “ms” latency claim is accurate; the SDK only states “low-latency” at kumoai/trainer/online_serving.py:59.
• The endpoint methods and optional real-time features shown in the deleted section otherwise match the SDK.

Local checks

$ git diff --check origin/main...HEAD
(no output; exit 0)

$ rg -n -i "online serving|online endpoint|launch_online_serving_endpoint|realtime" CLAUDE.md AGENTS.md README.md docs skills meta context
context/_gaps.yaml:167:      Health check for online serving endpoints. update() accepts

Grammar/spelling: No issues found in the changed Markdown.

[leena-kang]

Changes needed before merge

This is a follow-up review under the expanded non-agent Kumo-repo evidence policy. Recommendation: do not merge as-is. The authoritative evidence supports “gated/experimental,” not the blanket claim “unavailable to customers.” Keep the removal of the legacy snippet, but restore Online Serving as a qualified fine-tuned-SDK capability and point to the current provisioned workflow.

Evidence

• Product gating: kumo at fetched origin/master SHA 8e0685108b96a37fff2780b8a1491852deceae8e defaults ONLINE_SERVING_MODE to DISABLED (gira/controller/gira_controller.py:10-18), while the disabled route says the feature is Alpha and directs users to contact Kumo for early access (rest/public/routers/online_serving_jobs.py:163-169). This establishes limited rollout, not universal customer unavailability.
• Customer availability: docs at fetched origin/master SHA cfa17dd2e775b6341299da579a5a2180e2a85116 includes fine-tuning/online-serving in customer navigation (docs.json:166-175) and documents live endpoints plus tenant credentials provisioned by the customer’s Kumo team (fine-tuning/online-serving.mdx:7,20-28,97-114). Its history explicitly records the customer-facing intent: git -C ~/code/kumo/docs log origin/master --format='%H %s' -- fine-tuning/online-serving.mdx → 5aa4a285f7118670bb6a5df70f27113eabac7275 Simplify online serving page into a customer-facing quickstart (#122).
• Current SDK status: kumo-sdk at fetched origin/main SHA 0b992d46cbbd10a58e023f166d9a6da4361de40f exposes the separate deployment client (kumoai/online/__init__.py:1-4,66-79,228-273) and explicitly warns that it is experimental (kumoai/online/__init__.py:59-63). DistillationTrainer and export_model implement the documented preparation flow (kumoai/trainer/distilled_trainer.py:16-31; kumoai/artifact_export/job.py:133-145). Implementation alone would not prove entitlement, but together with the customer docs and early-access gate it does not support the PR’s all-customers rationale.
• Legacy snippet: the older TrainingJobResult.launch_online_serving_endpoint still exists (kumo-sdk, same SHA, kumoai/trainer/job.py:260-279), but its future blocks with result(), not the deleted attach() call (kumoai/trainer/online_serving.py:171-188). Removing that exact snippet is therefore reasonable; removing every capability/decision mention is not.

Required accuracy fix

Replace the blanket removal with concise guidance that Online Serving is experimental/provisioned and uses the current distillation → export → kumoai.online deployment flow. Do not restore the deleted attach() example or an unqualified millisecond guarantee.

Local checks

• git diff --check origin/main...HEAD — passed.
• Markdown fence-balance check on both changed files — passed; no YAML changed.
• No targeted runtime test: customer entitlement/rollout is not testable locally, and the authoritative source/config/docs evidence resolves the review question.

Grammar/spelling: no standalone issues found.

[9Olive]

Thanks @leena-kang — I agree the blanket removal overstated things, and I've revised the PR accordingly (pushed 7d0c6e7, merge conflicts with main resolved).

Verified your evidence against source: online serving is gated/experimental, not unavailable — ONLINE_SERVING_MODE defaults to DISABLED and the disabled route is Alpha/early-access (kumo), kumoai.online exists but emits an ExperimentalWarning (kumo-sdk kumoai/online/__init__.py:59-63), and the legacy TrainingJobResult.launch_online_serving_endpoint() future blocks with .result() — not .attach() (kumoai/trainer/online_serving.py:183), confirming the removed snippet was buggy.

So instead of deleting every mention:

• Kept the removal of the broken launch_online_serving_endpoint() / .attach() snippet.
• Restored online serving as a qualified experimental capability: a concise "## Online Serving (Experimental)" section in sdk-overview.md pointing to the current distillation → export_model → kumoai.online flow (no .attach(), no unqualified millisecond guarantee), and re-added the comparison-table row / prose in rfm-vs-training.md marked experimental + Kumo-provisioned.
• The merge also picks up main's kumoai.experimental.rfm → kumoai.rfm rename.

git diff --check origin/main...HEAD is clean.