[robdobsn]

Problem

PR #17 changed RaftFileHandler to carry on past an ack that covered only part of a batch, and to treat an ack arriving 700 ms or more after the last block as the device's timer ack.

Since RaftCore PR #14, the device acks the first block out of sequence at once and keeps its acks in step with the host's batches. With that firmware the PR #17 rule goes wrong whenever a block is lost mid-batch:

1. The immediate ack shows progress, for the blocks before the gap, so the handler carries on sending.
2. The device drops those blocks without acking them.
3. The handler then waits 1 s for the device's timer ack, once per lost block.

Change

Each batch starts from the latest ack again, as it did before PR #17. PR #17's two fixes stay:

• Acks that arrive while a batch is going out count. Before PR Feature/upload resend after loss #17 they were ignored, so after any loss every later batch waited for the timer ack.
• A late ack behind one already received is ignored.

The timing-based timer-ack detection is removed. At most one batch is in flight, as before PR #17. PR #17 allowed two.

Effect

Upload tests: a simulated device with RaftCore's ack rules, 200 blocks at 15 ms each (BLE-like), 6 lost blocks.

Firmware, lost blocks	Before PR #17	PR #17	This PR
Current (RaftCore #14), mid-batch	11.2 s, 5 timer acks	12.9 s, 6 timer acks	5.4 s, 0 timer acks, 239 blocks sent
Current, at the start of a batch	12.4 s, 6 timer acks	6.7 s	6.6 s, 251 blocks
Older, mid-batch	26.8 s, 21 timer acks	6.2 s, 309 blocks, 25 in flight	7.4 s, 391 blocks, 10 in flight
Older, at the start of a batch	29.2 s, 23 timer acks	6.0 s, 301 blocks, 20 in flight	7.0 s, 367 blocks, 10 in flight

With older firmware this PR is about 1 s slower than PR #17, but uses the original flow control and is still about 4× faster than before PR #17. RaftCLI's RICREST upload uses the same rule, and its own simulator gives the same block counts.

Tests

• RaftFileHandler.test.ts now covers current and older firmware, with blocks lost mid-batch and at the start of a batch, plus the last block lost and the late-ack case. It checks timer acks, blocks sent and blocks in flight.
• 3 of the 8 tests fail with PR Feature/upload resend after loss #17's code.
• All 237 tests pass; lint and build are clean.
• Not tested on a device.

🤖 Generated with Claude Code