v1.6.0 — Subagent transcript + tool content indexing

Retrieval coverage and correctness release. On a real ~10,000-file transcript corpus this took indexed documents from 31K to 677K and recall@5 on a 17-query set from 12/17 to 17/17 (recall@1 82.4% → 94.1%).

Coverage

• Nested subagent transcripts are now indexed. Enumeration was non-recursive, so only top-level session files were ever opened — 74% of transcript files were never read. Nested files get a disambiguated doc_id, because subagent transcripts carry their parent session id and would otherwise collide with and overwrite parent chunks.
• tool_use, tool_result and thinking blocks are now searchable. The extractor kept a content block only if it was type == "text" or carried a literal text key, dropping ~73% of blocks in a real transcript — so decisions made through tools were invisible to search. Prose is preserved ahead of tool content, and turn-pairs split rather than truncate, so adding tool output cannot evict the actual conversation.

Silent-failure fixes

Both of these reported success while doing the wrong thing, which is why they survived so long.

• Enumeration is fault-isolated. An unreadable or dangling path raised ENOENT during file collection — upstream of the existing per-file parse isolation — so one bad path out of thousands aborted the entire pass. Worse, the error was caught and converted into return 0, leaving the error list empty and the job reporting "complete" over a corpus it never touched. Bad paths are now skipped and counted, and genuine failures propagate into /index/status.
• EMBEDDING_MODEL is now reachable. The worker launcher overwrote the environment variable with a hardcoded constructor default, so configuring a different embedder silently had no effect.

Embedding configuration

• EMBEDDING_DEVICE — selects the embedding device, auto-detecting CUDA and falling back to CPU, and logs the resolved device at startup. ChromaDB's default is "cpu", so embedding silently stayed on CPU on GPU-capable hosts.
• EMBEDDING_QUERY_PREFIX — applies an asymmetric model's query-side instruction. ChromaDB applies one embedding function to documents and queries alike, but models such as arctic-embed, bge and e5 prefix only the query; without this, query and passage vectors land in different regions of the embedding space.
• UPSERT_BATCH_SIZE — the previous fixed batch of 20 fit inside a single sentence-transformers internal batch, starving a GPU.
• The unused bge-small generator no longer auto-selects CUDA, where it held ~390 MiB of VRAM without ever being used.

Ranking

• Weighted reciprocal-rank fusion with the vector arm down-weighted. Equal-weight fusion measurably regressed recall, and a unit test fails if the weights revert.

Upgrade notes

Indexing behaviour changed substantially — a full reindex is required to pick up subagent transcripts and tool content, and the resulting index will be considerably larger. Existing chunks update in place; the first part of each turn-pair keeps its original doc_id.

Tests: 26 passing.