lc cannot materialize a spec with sub-analyses

[EiffL]

lc materializes a flat analysis. A spec that declares analyses: is not buildable: ASTRA qualifies an output declared in a sub-analysis as <analysis>.<output>, and assets.output_path refuses an id carrying a dot, because the manifest sidecar (.<output_id>.manifest.json) is recovered by partitioning the filename on its first dot — a dotted id would name a manifest that could not be told from the payload, or that two outputs would share.

Why it is refused rather than supported

Supporting nesting means answering where a nested output's file goes, and every answer needs a decision that is not lc's alone to make:

• A scope directory — results/<universe>/<analysis>/<output>.<format>. Simple, one results tree.
• Beside the sub-analysis's own spec — an external (path:) sub-analysis is documented in astra-spec's llms.txt as "itself a valid ASTRA project with its own astra.yaml and optional universes/", which argues its results belong with it. That means several results trees, which reaches .gitattributes routing, the dirty-tree refusal's path split, and the crate's file walk.
• The qualified id as the filename — rejected: it makes the sidecar ambiguous, which is the silent-corruption case above.

There is also an open question about which universe a nested output is filed under when the parent universe selects one of the sub-analysis's own (analyses: {hod: {universe: fast}}) — two parent universes selecting the same one would otherwise collide.

Prerequisites

• astra-tools: validate_universe_file validates against the unresolved tree (it calls load_yaml, never resolve_analysis_tree), so a universe naming decisions inside a nested external sub-analysis fails with [UNKNOWN_ANALYSIS] even though the spec is valid. Depth 1 survives because the sub-analysis key exists unresolved; depth 2+ does not. Worth fixing upstream before lc builds on path:.
• astra-spec: no path: example exists anywhere in examples/ or tests/data/valid/; the only nested example, examples/iris_pipeline/, is inline. The convention for where sub-analysis directories live is unsettled — the specification shows stages/catalog_cleaning, while analyses/ is used in the docs for ref:-ed independent records.

Notes

Not a regression in behaviour anyone relied on, but worth stating plainly: before the one-file layout, a nested output landed in a flat directory named by its dotted id. See also #201.