[deepakgudiseva-nexthop]

Pre-submission checklist

• I've ran the linters locally and fixed lint errors related to the files I modified in this PR. You can install the linters by running pip install -r requirements-dev.txt && pre-commit install
• pre-commit run

Summary

At high ECMP width a route replace or withdraw makes the oper delta SwSwitch pushes to the HwSwitch large (a replaced 128-way route is ~110-220 KB), and that delta is copied several times on the way: into the per-matcher cache, when it is filtered per switch, and when it is handed to the HwSwitch. sw_agent's peak memory during large route updates scales with those copies; with a few thousand 128-way routes it can exceed the agent's memory limit.

This builds a sole consumer's delta without caching it and moves it to the HwSwitch, filters per-switch deltas in place, and moves rather than copies where the source is reset afterwards. No behaviour change.

Test Plan

• Agent unit tests pass with this change in our fork's CI; this commit is the same change applied to main.
• pre-commit run (clang-format 21.1.2) clean on the modified files.

[deepakgudiseva-nexthop]

Memory impact (#1906 + #1907)

Setup: 128-way ECMP, 22k routes (~2.8M paths), with enable_nexthop_id_manager, resolve_nexthops_from_id and drop_inline_nexthops_with_id enabled.

fboss_sw_agent memory, GiB:

	drop inline only	+ #1906 + #1907
Peak, route learn	5.8	2.2
Peak, route withdraw	5.6	2.1
Steady state	0.48	0.46

Together the two PRs cut sw_agent's peak during route updates by about 3.6 GiB (~60%). Steady state stays the same.

[deepakgudiseva-nexthop]

/autorebase