Index tool_use inputs, so a session is findable by the file path and the command

[devcodes9]

Goal: make a session findable by the file path and the command, not only by their output.

_flatten_content indexes text and tool_result, and drops tool_use inputs. So you can
find a session by what a command printed, but not by the command. Same for a file: by its
contents, not its path.

Fix: index the identifying fields of a tool_use input (file_path, command, pattern,
query, url). About 3.2% of transcript bytes.

Already measured, do not redo

Numbers from a ~740-session corpus, freshly built index, so compare arms with each other and
not with the published gold and held-out figures.

• Tool output is 71.6% of the index. Bash is 55.5% of it, Read 37.8%, WebSearch 1.5%.
• Keep tool output. Dropping it takes tool-content recall to zero (@1 0.283 to 0.000).
• No blanket cap. Capping at 1500 saves 24% of index size and costs 12% of tool-content
  recall. Not worth it.
• No per-tool caps. Bash and Read are the biggest, and also the most useful: test
  failures, stack traces, and the session that touched a symbol.
• Snippets are fine. 88% of displayed evidence lines already come from typed text.

Bigger win, separate

Per-keystroke cost is mostly reading sessions.tsv (0.091s of 0.246s) and _at_word_start.
Reading the TSV as bytes and counting with a compiled \b regex was estimated at 0.33s to
0.10s, with no change to ranking.