docs: correct env claims, the Postman full tier, and Quick Start model examples

[jsugg]

Three documented facts that were not facts. Each verified against source or the
live API, not assumed.

RATE_LIMIT_REDIS_TOPOLOGY

The doc asserted the Render deployment "keeps RATE_LIMIT_REDIS_TOPOLOGY=external".
The variable is not set on the service at all (confirmed against the live
API). The behaviour is right — external is the code default
(config/rateLimitStore.js:51,61) — but stating it as configuration sends
anyone auditing the dashboard hunting for a variable that was never there.

Fixed in the repository rather than by setting the variable on the service: the
deployment already behaves as documented, so there is nothing to change in
production, only something to stop mis-stating.

SCRAPER_REQUEST_TIMEOUT_MS

The documented 10000 default is correct (config/index.js:110). The service
runs 90000, and the doc never mentioned it. The deviation is deliberate —
hosted vision models routinely exceed 10s per image, and the free-tier instance
is slower still — so it is recorded with its reason, rather than looking like
drift to whoever finds it next.

The Postman full tier

Described as "Full deterministic harness for main and production". Wrong
on both counts:

• npm run postman:full appears in exactly one workflow,
  local-provider-integration.yml, whose triggers are workflow_dispatch, a
  weekly schedule, and paths-filtered push/PR to main. There is no
  production trigger.
• CI does run on production pushes, but hardcodes NEWMAN_MODE: smoke with
  no branch conditional. production gets smoke, never full.

The sibling tiers are written as trigger policy (smoke = "required on pull
requests", live = "Manual or scheduled"), so this bullet read as trigger
policy too — and promised production coverage that does not exist. It now
separates capability from execution policy.

Quick Start model=replicate

Not a style nit. Providers register only when their env keys are set, and an
unregistered model throws Unknown model → 400 UNKNOWN_MODEL /
unsupported_value (descriptionController.js:215-221). An install configured
for any provider other than Replicate — an OpenAI-only setup, say — got a 400
following Quick Start verbatim, with nothing in the surrounding text to explain
why.

The examples now use <configured-model> and say the value must name a
provider configured in .env.

Verification

docs:validate 8 files / 0 violations · docs:lint 0 errors · no
model=replicate remains in the README.