Selective contribution to the built-in `context` target raises `'context' is undefined` — dict form routes every target to the named-aggregator lookup

[Steve-Allison]

Describe the bug

The schema allows a ContributeElement to be dict[str, ContributeValue], documented as
"Contribution of a specific value instead of the default one", and ContributeTarget includes
context. But contributing a specific value to the built-in context target fails at runtime:

text:
- def: STEP
  text: "FULL-RESULT-BODY"
  contribute:
    - result
    - context:
        value: "TINY-SUMMARY"
- "end"

PDLRuntimeError: Error during the evaluation of ${ context }: 'context' is undefined

The failure occurs during contribution processing itself — no downstream consumer of the context is
involved. The equivalent named-aggregator form (an aggregator: declared in defs, as in
examples/aggregators/contribute_in_file.pdl) works correctly.

Root cause (from reading the interpreter)

In process_contribution, the plain-string cases result / context return early, but the
dict() case has no special-casing for the built-in targets: after evaluating the value it falls
through to get_contribute_aggregator(block, target, scope, loc), whose str() branch does
get_var(target, scope, loc) and requires the result to be an Aggregator. context is not a
scope variable, so the lookup raises; if a variable named context did exist, it would then fail
"An aggregator was expected". The same applies to result as a dict target.

• process_contribution (main @ 045f052):

  prompt-declaration-language/src/pdl/pdl_interpreter.py

  Line 1637 in 045f052

  def process_contribution(

• get_contribute_aggregator (main @ 045f052):

  prompt-declaration-language/src/pdl/pdl_interpreter.py

  Line 2803 in 045f052

  def get_contribute_aggregator(

Expected behavior

contribute: [result, {context: {value: <expr>}}] contributes the evaluated value to the
background context in place of the block's result — the "small summary to context, full result
returned" pattern the schema description implies — mirroring how the plain context target and the
named-aggregator dict form already behave.

Versions / reproduction

• Reproduces on prompt-declaration-language 0.9.3 (latest PyPI release) and on current main
  (045f052 — the dict path is unchanged by refactor: better separate leaf and structured blocks #1669).
• Zero-cost repro available: the YAML above with any model-free block, or with
  parameters: {mock_response: ...} on a model block.
• Possibly related: Support for user-defined contexts #437 (user-defined contexts) generalises contribution targets but does not
  mention this defect; fix: contribute directly a value to the context #991 fixed the wrapping of successfully contributed values, not the
  dict-target routing.

Why it matters

This is the language's only mechanism for contributing a compressed representation to the
accumulated context while returning the full result — the natural tool for large-payload pipelines.
As it stands, the schema advertises the capability and the interpreter cannot execute it.