NVIDIA-NeMo#4216)

## Problem

FastAPI 0.142.0 to 0.142.4 ([release
notes](https://github.com/fastapi/fastapi/releases/tag/0.142.0),
[fastapi/fastapi#16403](fastapi/fastapi#16403))
configure OpenTelemetry export by themselves at ASGI startup whenever
`OTEL_EXPORTER_OTLP_ENDPOINT` is set. Gym sets that variable for every
server process so that nemo-lens finds the collector, and nemo-lens has
already registered providers with their own exporters. FastAPI attaches
a second `BatchSpanProcessor` and a second
`PeriodicExportingMetricReader` to those providers. Gym does not pin
FastAPI, so every job environment resolved since 2026-09-29 picked this
up.

Observed in every gym-native run since then:

- every span is exported twice: each trace appears twice in Tempo and
the collector's span-metrics counts are doubled;
- every metric is exported by two readers with separate aggregation
state, so the cumulative `_sum` of the Gym histograms carries last-digit
dips that Prometheus treats as counter resets; `increase()`/`rate()` on
a sum then adds the whole running value (mean rollout duration showed
hours instead of minutes, 60 of 60 runs affected).

## Change

FastAPI 0.143.0
([fastapi/fastapi#16476](fastapi/fastapi#16476),
from [discussion
#16473](fastapi/fastapi#16473)) makes the
automatic exporter setup opt-in, so only the 0.142 series
double-exports. This PR excludes that series in `pyproject.toml`
(`fastapi!=0.142.*`) and refreshes `uv.lock`; the locked version stays
0.135.2. An earlier revision wrapped the six `FastAPI()` sites in a
factory that switched the native telemetry off; the exclusion replaces
it because it also covers the bare `FastAPI()` apps outside the server
bases and needs nothing from future servers.

Environments already built with FastAPI 0.142 keep exporting twice until
they are reinstalled.

## Verification

A gym-native run with the factory revision as the job-side Gym
(terminal-bench miniswe, 8 rollouts, 1 node) showed one
`gym.rollout`/`gym.verify` span per rollout, one metric datapoint per 10
s with consistent sums, and
`resets(gym_rollout_duration_ms_milliseconds_sum[1d]) = 0`, against 2
spans per rollout and up to 684 resets per run on `main`. The exclusion
produces the same single export by resolving FastAPI to 0.143.0 instead
of switching the feature off.

Signed-off-by: Ingo Guehring <iguehring@nvidia.com>
Co-authored-by: Claude Fable 5.1 <noreply@anthropic.com>