[renatex314]

Point 1 of #10, as agreed in the discussion there: one report the server refuses no longer has to cost the rest of its batch. Opt-in, so existing ^0.9.0 clients see no change. Independent of #11; the two merge cleanly, and the merged result passes everything (see Testing). Intended for 0.10.0 with #11.

Problem

A batch is all-or-nothing. One refused report (an unknown value, a full capacity dimension, oversized props, half a position) fails the request with a 400 naming only that report, and nothing in the batch lands. The Node client sends up to 1000 reports per request, so one bad device costs up to 999 good ones.

It's worse in the Reporter: it puts a failed batch back in its queue, and a report refused on its merits is refused every time. So one bad device makes every later flush fail and blocks every device queued with it.

Server: POST /positions?partial=true

• Each report is applied on its own. The good ones land, and the refused ones come back in rejected, in batch order:

  {"accepted":2,"stale":0,"devices":2,
   "rejected":[{"index":1,"id":"truck-9","code":"bad_request","error":"device \"truck-9\": unknown value \"idel\" ..."}]}

  code and error are exactly what the 400 would have carried. id is null for a GeoJSON Feature that carried no id.
• accepted + stale + patched + unknown + rejected = batch size.
• rejected is present, possibly empty, whenever the flag is set. That's how a client tells a server that honoured it from one that predates it.
• Still request-level, because none of them can be pinned on one report: a body that doesn't parse, an oversized batch, an overloaded server, and ?partial= set to anything other than true/false.

Without the flag: a batch is still all-or-nothing, and a batch with one bad report gets the same 400, message and code as before. The existing tests that pin that behaviour pass unchanged.

Change in both modes: reports are now checked before any value is resolved, in this order: position/geometry/id, then props, then values. Resolving a value on a capacity dimension takes a slot permanently, and a report refused for its props used to take one anyway. Covered by a_refused_device_takes_no_capacity_slot.

One visible side effect: when a batch has several bad reports of different kinds, the 400 may name a different one than before. Example: a bad value at item 1 and oversized props at item 2 used to name item 1, and now name item 2. (The commit message says the first-named error is unchanged too; that's wrong in this case.)

Node client

• partial: true option: available on report, patch, reportGeoJSON and reporter. The default stays false.
• Results: they carry rejected, with index renumbered from each chunk to the caller's whole list.
• Acknowledgement check: checkAck counts rejected.
• Reporter:
  • A refused report is dropped instead of requeued, and counted in stats.rejected.
  • It's passed to onError as a NetClusterError with body: { code: 'rejected', rejected }.
  • The flush itself succeeds, because the rest of the batch landed.
• Upgrade order: against an older server the flag is ignored and a bad chunk is refused whole, as before. So the client can be upgraded before the servers.
• Docs: index.d.ts has the new types (WriteOptions, RejectedReport), and both READMEs describe the behaviour.

Testing

• Server: new tests/partial.rs (8 tests), covering:
  • Without the flag, the same first error and code, and nothing applied.
  • With it, good reports land and bad ones are named in batch order.
  • Accounting alongside stale/patched/unknown.
  • A full capacity rejecting only the overflowing device.
  • A refused report taking no slot.
  • GeoJSON: null geometry, missing id, swapped coordinates, unreadable properties and an unknown value.
  • A batch with nothing good.
  • The flag being strict and always answered.
  • Each case also checks index invariants with verify().
• Client:
  • A unit test for renumbering across chunks and the acknowledgement check.
  • Integration tests against the real binary for report, patch, reportGeoJSON, and a Reporter that drops the refused report and doesn't resend it.
• CI steps run locally, in rust:1.98-slim-bookworm:
  • cargo fmt --check and cargo clippy --all-targets --all-features with -D warnings: clean.
  • cargo test --release: 207 pass.
  • Node npm test: 19 unit, 44 integration and 29 CLI tests pass.
• Merged with Let a dimension hold no value instead of refusing it or filing it at 0 #11: formatting and clippy are clean, 210 Rust tests pass, and the Node tests pass.

🤖 Generated with Claude Code