Phase 1 of the catalogue repair: the validator, not the examples.

orchestrator validate rejected 108 of 117 examples. Only 23 actually fail
to compile when run. The gate was measuring the validator's strictness rather
than whether a pipeline works.

The decisive case — one pipeline, three ways to name a parameter, all three
rendering identically at run time:

Four causes

Three environments, three filter sets. Validator 56, compiler 61, runtime
70 — each built its own. {{ title | slugify }} was "Unknown filter" in a
pipeline that renders it fine, and fixing only the validator moved the failure
to truncate_words in the compiler. All three now come from
create_pipeline_environment, sourced from the runtime's registry.

parameters. / inputs. unrecognised. The data-flow validator read
parameters as a task id and reported an undefined task by that name.

A declared parameter without a default was reported undefined. It is
declared; only its value is deferred to -i name=value. A pipeline was told
Undefined variable: 'output_path' about the name it had just declared.

A hardcoded list of twelve output field names. Anything else was "does not
produce output" — asserting knowledge the validator does not have about a model
or <AUTO> step. Now an error only when the task declared its outputs, a
warning otherwise, and the warning still records the data-flow edge.

Results

My prediction was wrong, by a lot

I estimated Phase 1 would unblock ~85 files. It unblocked 9.

The inference was bad: "85 examples compile at run time" does not mean
validate has only one objection to each. Most files carry several independent
error classes, so removing one class halves the error volume without moving
the file count. I conflated two different measurements.

Mutation testing

Each fix reverted against the new tests:

The last one survived at first. I had changed the severity and the
message, and my test asserted the old message string was absent — which the
reverted path no longer emitted either. It tested wording, not behaviour. The
test now asserts the pipeline compiles.

What remains is real, and I checked rather than assumed

• output_path (92 occurrences, 16 files) is genuinely undeclared. Those
  pipelines never declare it and the runtime fails identically:
  'output_path' is undefined. The validator is correct.
• execution (97) — the runtime does not populate it either, so
  data_flow_validator.py:485 accepting it is the bug: too permissive. Fixing
  that makes more things fail, which is why it is deliberately not in this PR.
• Loop variables (item, loop, row, col_data) unclassified.
• 17 files missing tools, 13 schema, 6 YAML syntax, 6 model, 4 file-inclusion.

Nothing under examples/ is modified or deleted. Full notes in
notes/example-catalog-diagnosis.md.