feat!: signal-driven conditional pipeline with single full/smoke config

[sousekd]

Replace the four shipped pipelines and their ad-hoc knobs with one
signal-driven full pipeline (plus a minimal smoke), and make both external
source providers pass their options through opaquely.

Conditional execution

• Add a Condition contract (signal-name leaf plus all/any/not) and an evaluator;
  a step runs iff (runIf absent or true) AND (skipIf absent or false).
• Orchestrator gates each step against the live signal bag; gated steps are
  recorded as skipped (run_if_unmet / skip_if_met), not executed.
• Step schema gains runIf, skipIf, and enabled.

classify-url step

• Add a built-in step that runs once and emits boolean signals from URL rules.
• Each rule carries one signal and one matcher: extensionIn, pattern, or
  anyHost; matchers are precompiled at parse time.
• Shipped rules emit binary_doc (PDF/Office, arXiv /pdf/) and code_host.

Single pipeline (breaking)

• Collapse truncate / clean-deterministic / clean-llm / clean-combined into one
  full pipeline toggled by *_ENABLED env, plus a small smoke pipeline.
• Source routing is explicit and gated: fetch_docling (runIf binary_doc) ->
  fetch_firecrawl -> fetch_http fallback; readability skipped on code_host.
• enabled:false steps are dropped at compile time, so a disabled step's
  provider is never required at startup.
• DEFAULT_PIPELINE default truncate -> full.

Opaque provider options (breaking)

• Both providers keep output typed and move every other knob into an opaque
  options record (inline YAML object or a JSON-string env blob).
• Docling: doOcr / tableMode -> DOCLING_OPTIONS; rename docling-ocr ->
  docling-default.
• Firecrawl: onlyMainContent / stripBase64Images / parsePdf -> FIRECRAWL_OPTIONS;
  raw-PDF passthrough now detected by the %PDF magic header, not a flag.

Breaking

• Pipelines renamed; DEFAULT_PIPELINE default is now full.
• Provider docling-ocr -> docling-default.
• Env: FIRECRAWL_PARSE_PDF -> FIRECRAWL_OPTIONS; new DOCLING_OPTIONS;
  SOURCE_PROVIDER removed; per-provider timeouts; CLEAN_* / SUMMARIZE_* ->
  LLM_CLEAN_* / LLM_SUMMARIZE_*; new *_ENABLED toggles.

Docs, env examples, and Compose files updated to match.