schema.py: dedupe cyclopts help= vs Field(description=) across config fields

[arekay-nv]

Follow-up from PR #406 review: #406 (comment)

Background

Several config fields in src/inference_endpoint/config/schema.py set the same help text twice — once as cyclopts.Parameter(help=...) and once as Field(description=...). This is redundant:

• cyclopts (4.10.0) reads Field(description=...) natively for --help. When both are set, help= wins; delete help= and cyclopts falls back to description → identical CLI output.
• scripts/regenerate_templates.py (_collect_comments._help) already reads description first, help only as fallback.

So Field(description=...) can be the single source of truth for both the CLI --help text and the YAML template comments.

Proposal

Make Field(description=...) canonical and drop the redundant help=. Keep cyclopts.Parameter(...) only where it carries CLI-only metadata (alias, negative, name, show, required) — those have no Pydantic equivalent. No decorator/machinery needed: cyclopts already does the injection, so this is less machinery, not more.

(Avoid the bare attribute-docstring form — pydantic doesn't put it in FieldInfo.description, so the template generator would lose the comment.)

Scope (as of HEAD 6d28c00)

16 fields set both help + description → 10 identical (safe to collapse), 6 silently divergent; plus 17 help-only and 50 description-only (already single-source).

Identical — just drop help=

• ModelParams.seed
• WarmupConfig.{enabled, n_requests, drain, warmup_random_seed} — salt already collapsed in fix: Salt failure is a hard error #406
• DrainConfig.{warmup_timeout_s, performance_timeout_s, accuracy_timeout_s}
• ProfilingConfig.engine

⚠️ Silently divergent — CLI --help and the YAML comment already say different things

Single-sourcing prevents exactly this drift; decide per field which text is correct before collapsing.

field	help= (CLI)	Field(description=) (YAML/schema)
EndpointConfig.endpoints	Endpoint URL(s)	Endpoint URL(s). Must include scheme, e.g. 'http://host:port'.
EarlyStoppingConfig.enabled	Report MLPerf early-stopping percentile estimates for TTFT/TPOT/latency	Early-stopping percentile estimates (default on)
Settings.service_ready_timeout_s	Seconds to wait for metrics/event-logger services to start	Seconds to wait for metrics-aggregator/event-logger services to become ready.
ProfilingConfig.urls	Override URL(s) for profiler triggers; defaults to endpoint_config.endpoints	URL(s) the profiler start/stop triggers are derived from. When None, derived from endpoint_config.endpoints instead…
DrainConfig.metrics_drain_timeout_s	…Set to 0 to wait indefinitely. Increase for very large datasets…	…after ENDED (default: 0 = unlimited). An incomplete drain is surfaced via n_pending_tasks > 0, never silently dropped.
DrainConfig.metrics_tokenizer_workers	…0 defers all tokenization to the end-of-run drain, which always uses the auto-sized sharded pool.	…(default: 2; 0 = defer everything to the end-of-run drain).

Already done

PR #406 collapsed only WarmupConfig.salt (that PR edited its help text in both places, so it was in scope). This issue tracks the remaining 15 fields.