prospect diff: compare nodes across versions, runs, and points in time

[elecnix]

Builds on #50 (as-of reads). Part of the point-in-time set with #51 and the retraction issue.

Why a diff here is stronger than a text diff

Content-addressing makes a difference attributable, not merely visible:

observation	meaning
same input_key, different output_key	identical recipe — the model reached a different conclusion
different input_key	the recipe changed; compare its components to say which

And the components are already separated: analyzer version, config fingerprint, source-set hash. So a diff can state precisely why two nodes differ — the author bumped the analyzer, the user changed config or model, or the inputs moved. A text diff cannot distinguish those, and that distinction is usually the whole question.

Proposed change

prospect diff --unit <analyzer> <source_set_hash>   # across the revises chain
prospect diff --runs <A> <B>                        # what two runs concluded differently
prospect diff --as-of <T1> <T2>                     # added / removed / changed by analyzer

Node content is canonical JSON, so comparison should be structural (changed fields) rather than line-based, with a summary line per node and details on request.

Motivating uses, all encountered in practice

• "How did our conclusion about this turn change when the prompt improved?" The revises chain already links versions of one logical unit; nothing surfaces the content delta. Answering it currently means reading content_json by hand.
• Model and prompt bake-offs. Because the resolved model is part of the config fingerprint, different models' verdicts on the same term already coexist as separate nodes. Comparing them is the natural read — and it was recently done by hand-writing a throwaway SQL harness, when --runs A B would have been a first-class command. The same applies to comparing prompt versions: v1.1 vs v1.2 verdicts sit side by side already.
• Regression triage. When proposal quality moves, --as-of localises when, and --unit shows what actually changed in the conclusion.

Notes

• --runs needs run ids to be discoverable; prospect analyzers or a prospect runs listing should show recent runs with their mode, model spec, and node counts.
• For analyzers with many nodes per session, default to a summary (counts of added / removed / changed per analyzer) and require a selector for full content diffs.
• Diffing should read through the same live/as-of filters as prospect: --as-of reads for querying the graph at a point in time #50 and the retraction issue, so a retracted node appears as removed at T, not as though it never existed.

[elecnix]

Landed and merged: prospect diff compares nodes across versions (--unit), runs (--runs), and points in time (--as-of), with attributable content-addressed deltas.