Addresses findings 1 and 2 of the review of #448.

The shared registry was not shared

#448 routed the validator, compiler and runtime through one filter registry and
asserted it by comparing filter names. The compiler called its own
_register_custom_filters() on the very next line, replacing eleven of those
implementations. Name sets matched, so the drift test passed. Semantics did not:

The compiler's default treated every falsy value as absent and raised on an
undefined one — the single case the filter exists for. to_json, slugify and
regex_search diverged too, which the review had not caught.

All eleven filters the compiler registered were already in the runtime registry
and none were unique to it, so the 110-line method and its call site are
deleted rather than reconciled.

Behaviour, not registries

test_no_filter_diverges_between_environments renders every filter in the
registry through all three environments against probes chosen for where
implementations part company — undefined, None, empty string, zero, false,
non-ASCII, non-string — and compares results.

A hand-written table covers default's undefined-vs-falsy distinction,
malformed JSON, regex, paths and unicode explicitly. Three cases run end to end
through the CLI, so the assertion reaches a real file rather than stopping at an
environment.

Two comparison artefacts I had to handle rather than ignore: lazy filters render
as <generator object ... at 0x...> whose address differs between any two
renders, and random/shuffle are not functions of their input. Addresses are
masked; the two nondeterministic filters are excluded by name.

A test of mine that could not fail

test_a_pipeline_using_a_runtime_filter_validates accepted any error whose text
lacked the word "filter", and fed from_json the literal "Hello World Report"
— not JSON. An unrelated rendering error satisfied the assertion. It now asserts
the pipeline compiles, on valid JSON. That is the second test in two PRs where I
asserted on error wording instead of behaviour.

Evidence

• Blocking suite 573 → 602 passed, 0 failed
• Catalogue unchanged at 18/117 — this PR fixes semantics, not acceptance
• ruff (CI rule set) and compileall clean

Not in this PR

Deliberately, so each stays reviewable:

• Finding 3, runtime namespaces. execution, pipeline, context, env
  are still accepted unconditionally and at least execution is never populated.
  Fixing it makes more examples fail, and it wants a typed context schema that
  validation and runtime population both derive from.
• Finding 4, the four integration cases for unknown output fields (warning
  reaching consumers, dependency affecting scheduling, declared schema still
  rejecting misspellings, nested fields).
• Findings 5 and 6, catalogue triage and the CI dashboard.