[ISSUE #7716] Normalize instance information across reporting paths. - #7785

Approving. The two defects are real and each is fixed at its own layer; I traced the whole write/read path to check that the string contract still holds.

Evidence for the two symptoms

HTTP path — HttpSyncDataService:219 sends SystemInfoUtils.getSystemInfo() as a raw header, while HttpLongPollingDataChangedListener:146 reads it with defaultString(..., "") and :152 re-encoded it with toJson(...). The dashboard's single parse therefore produced a String, not an object. Passing the header text through unchanged is the correct fix.

WebSocket path — ShenyuWebsocketClient:426-433 wraps the same payload into Map.of(BOOTSTRAP_INSTANCE_INFO, info) and serializes the wrapper, so the admin receives a JSON object whose value is object text. convertToMap then decoded recursively, which is why {"note":"{\"nested\":true}"} came out mutated and why malformed optional metadata could throw. Extracting the element from the outer tree (isJsonPrimitive() && isString() ? getAsString() : toString()) keeps the outer encoding while leaving nested values alone — that is the right read point.

Write / read convergence

onInstanceInfoReport (InstanceCheckService:196+) builds the DTO and calls handleBeatInfo, so the normalization at handleBeatInfo is on the path into instanceHealthBeatInfo, and syncDB -> InstanceInfoServiceImpl#createOrUpdate persists that normalized value. Reading normalizes again in buildInstanceInfoVO:114, which covers findById, list (also used by fetchInstanceData) and the paged query, and it does so without writing back — the mapper test that asserts the stored row stays double-quoted while findById returns the decoded text is exactly the assertion I wanted to see. normalize decoding at most one legacy layer, and re-entering it being idempotent, is what keeps this from becoming an open-ended decode chain.

Non-blocking

1. normalize returns element.toString(), so a stored/reported value with whitespace or unusual key order is compacted on the display path. Harmless, but it means normalize is not a string-level identity for pretty-printed input; the tests only feed compact text. A one-line comment on the canonical-form expectation would prevent a future reader from treating it as a no-op.
2. handleBeatInfo normalizes only in the else branch (first registration); an instance already in instanceHealthBeatInfo keeps the value it was created with and syncDB writes it back unchanged. That is pre-existing behaviour (metadata is never refreshed), but combined with this change it means a long-running admin converges only after a restart. Worth one line in the description, because the issue is about what the dashboard shows.
3. WebsocketCollector#onMessage now calls GsonUtils.getInstance().fromJson(message, JsonObject.class). A message that contains the marker but is not a JSON object throws JsonSyntaxException out of @OnMessage; master behaved the same via convertToMap, so it is not a regression — but a try/catch that logs and returns would be strictly better for a websocket entry point, and it is one line.
4. The HTTP path with the header absent is only covered implicitly. The parameterized websocket test covers null/empty, so a direct case on doLongPolling without the header, asserting the event carries "" and the heartbeat still lands, would close it.

CI

gh pr checks reported nothing on f375e0be; the runs were action_required (first-time contributor), so I approved the ci workflow run. The change spans 11 files and rewrites JSON semantics at three boundaries, so I am not recording an approval before pr_build is green on this head — I will flip it as soon as it is.

Follow-up: pr_build is green on the current head (3m9s), so the CI evidence I was waiting for now exists. Everything in my previous review stands — flipping to an approval.

Approved.

Good normalization work across all instance-info reporting paths. The HTTP long-polling listener no longer double-encodes the header JSON (the bug that wrapped object text in an extra string); the WebSocket collector reads only the outer frame and preserves inner JSON strings instead of re-serializing; and the new InstanceInfoUtils.normalize() canonicalizes legacy/over-encoded rows for display without rewriting stored values. Tests cover legacy compat, malformed-frame resilience, and the nested-string case (verified that double-wrapping collapses to {} as intended).

Minor note: CI checks have not reported on this branch yet; please confirm the workflow passes before merging.