plugin: hooks find the graph from the path a tool touched (#1453)

[swapnilpaliwal-sd]

Fixes #1453

What was wrong

Every hook (orient, direct, enrich, changes, refresh) looked for .axiomcode/out/graph.sqlite in the session's working directory only. Agents rarely work there. They start in a directory with no graph and read, grep and edit files in indexed trees by absolute path. Measured over 190 agents: 2,115 graph calls, and the hooks injected 3 times in total. One was a directive fired by a cat of a Markdown file. The other two were orientation loads. There were 0 Read or Grep enrichments. 308 of 309 sessions had a working directory with no graph.

Two smaller defects in the same place:

• direct.py filtered a Read by source extension but not a shell read, so cat NOTES.md got the directive.
• orient.py counted no .cs files (csharp: the prompt hook never says a C#-only tree has no graph yet, because orient.py's extension list omits .cs #1453). It also told a workspace directory holding an indexed project that it had no graph.

The change

New hooks/_where.py finds the repository for each event:

• It takes the path the tool touched: Read/Edit/Write file_path, Grep/Glob path, the graph tools' repo, and for a shell command its last cd and path arguments that exist.
• It walks up from the real path to the nearest main graph or per-language graph (.axiomcode/lang/<lang>/out/graph.sqlite).
• It falls back to the working directory only when the tool named no path.

A path with no graph above it stays silent, even when the working directory has one, because a search of another tree says nothing about this graph. A symlink into an indexed tree answers from the real graph. A symlink out of it stays silent.

Each hook uses it:

• The state, budget, stamp and log files live in the repository's .axiomcode, so they are kept per repository.
• A file whose language has its own graph is answered from that graph.
• refresh.py also refreshes the repositories the session worked in on events that name no path.
• orient.py also finds the graph from an absolute path in the prompt.

Found roots are cached per directory for the session. "No graph" is not cached, so a graph built mid-session is seen on the next read. An uncached walk is under 1 ms.

Tests

tests/hooks_from_path.py: 26 checks, all passing. The layout is a workspace with no graph, an indexed project inside it, an unindexed copy, and symlinks.

• Checked: Read, Grep with path=, sed -n by absolute path, cd <project> && grep, and a relative path.
• Controls: a tree with no graph, a path above the indexed root, a symlink out of the tree, cat of a non-source file, and a 5-file C# tree.
• Also passing: directive.py, enrich_budget.py, enrich_lines.py, mcp_first.py, hosts.py, no_symlink.py, and refresh.py for Python, Java and C#.
• python_names.py has one failure. It fails the same way on the unchanged base.

Smoke (Python, Java and C# project copies, fresh index)

Setup: a headless session whose working directory has no graph. The task was to Read one file in an indexed copy by absolute path, then Grep a name there. Before is the installed plugin; after adds this branch with --plugin-dir.

project	hook injections before	after
Python library, ~50 files	0	3 (directive, Read block, Grep preview)
Java web app, small	0	3 (Read marks the handler entry (http))
C# library, ~20 files	0	3

Hook latency (median of 20 runs, Read enrichment): 118 ms before with the working directory in the repo. After: 105 ms with it in the repo, 109 ms with it outside. The path lookup is 0.2 ms.

Deep scan deferred: the corpus-wide regression gate was not run for this change.