plan.py: check that a pattern's indices agree before planning across them

[litkhai]

Closes #40.

plan.py plans across an index pattern, mapping_to_ddl.py reads one index and run.py
loads into one table. So a pattern whose indices disagree lands two shapes in one table,
or fails on the first chunk from the odd index out — and a rollover alias with months of
backing indices that each grew their own fields is the normal Elastic case, not the exotic
one.

One _field_caps request over the pattern answers it, and the two findings get different
treatment because they are different problems:

Finding	Treatment
a field with two types across the pattern	refused. One ClickHouse column cannot hold a keyword and a long, so there is no plan to make. Narrow the pattern, or pass --allow-mapping-conflicts once the column has been decided
a field only some indices have	a warning. That column is empty for rows from the indices that lack it — usually a mapping that grew over time, occasionally the sign that this pattern is two datasets

Both lists land in plan.json either way, so the decision stays visible after the terminal
has scrolled.

mapping_to_ddl.py made the same distinction badly — a pattern produced not found in _mapping response (got keys: [...]). It now compares the resolved mappings: identical ones
are one shape and it says which index it read, different ones are refused with a pointer at
plan.py, which names the fields that disagree.

Building the fixture found the case worth having

With dynamic mapping on, indexing a string into an index that never declared the field
creates it as text there and keyword elsewhere. So "a field some indices lack" and "a
field with two types" are the same event in practice — my first fixture produced both
findings for one field, and testing them apart needed dynamic: strict. That is not an
artifact of the test; it is how a real pattern acquires conflicts.

Verified on

Elasticsearch 8.17.0, four fixture indices with dynamic: strict.

Checked	Result
agreeing indices	plans silently, no warnings
a field only one index has	WARNING, exit 0, recorded in plan.json as mapping_partial_fields
status as long on one and keyword on the other	refused, exit 1, both types and their indices named, with the two ways forward
--allow-mapping-conflicts	plans, exit 0, conflict kept in plan.json
mapping_to_ddl.py on identical mappings	note: … resolved to 2 indices with identical mappings; reading mtest-a, DDL emitted
mapping_to_ddl.py on differing mappings	refused, pointing at plan.py
single-index regression	full pipeline: 4 chunks, 300,000 rows, parity passing

AGENTS.md's scope entry moves from "not covered" to "detected, and still does not merge
two shapes" — and the failure table gains the refusal, with the note that
--allow-mapping-conflicts is for after the decision, not for getting past the message.

🤖 Generated with Claude Code