path '*': the methods behind a [by name] caller are counted, and bound: lines survive to the front door

[swapnilpaliwal-sd]

What

Two gaps in path '*' <target>, both from the week-2 external usage report (finding 8 / ask 2):

1. The closure stopped silently behind a by-name caller. A method that reaches a [by name] caller through resolved edges reaches the target too whenever the by-name site is real — on "only X can trigger this" (the report's Q3, their spec-writing shape) exactly the triggers the answer dropped under a verified: line. One reverse walk from the by-name caller set now counts them, with up-to-3 nearest NAMED examples (placeholders like <arrow> are pushed behind named methods):

   [by name] cancel_endpoint   api.py:5   — calls `cancel` (receiver not typed)
   behind them: 1 more method(s) in 1 file(s) reach these by-name caller(s) through the graph's edges — callers of the target too if a by-name site is really `OrderService.cancel`; nearest: outer_handler (api.py:13)
   

2. The front door dropped every bound: line. ax_blocks.render kept only run:/verified foot lines, so the honesty lines the report explicitly asks us not to tidy away were invisible on the default surface (CLI and MCP). They are kept now; run: stays last.

Evidence

• Both languages reproduce the gap (python + typescript 4-function fixtures; the resolved caller of the by-name caller was absent and uncounted in both).
• Scale + cost on a real graph (this repository, ~2,100 TS/JS files): a hot name (has) counts 907 behind-methods in 126 files, ~1s added; a mid-frequency name (6 sites) counts 8 in 1 file. Mid-size target before the walk: 2.7s; after: unchanged to the tenth.
• Controls: a target with only resolved callers prints nothing new; a caller of a method whose by-name site names ANOTHER method is not counted (suite check).

Tests

• Extended python/path-crosses-framework-and-byname (java: path '*' X leaves out a caller on an untyped receiver that impact lists as [by name] #1421's case): the behind count, plus the keyed negative control. 9/9.
• Full suite legs: python 289/289, typescript 225/225, java 329/329, javascript 285/285, csharp 207/207; front_door.py 17/17 (the "answer ends with run:" invariant kept by ordering, not by dropping the bound).

[swapnilpaliwal-sd]

Two more commits from the AllScale pilot feedback:

• path --json carries what the prose says (their Finding 1): verified is true/false once hops were checked (null stays "nothing was checked"); unresolved_inside / unresolved_closure are counted numbers where 0 is a counted zero; bound is the same text the prose prints. Suite checks compare both formats on one query, at zero and at a nonzero count.
• context is back on the MCP surface: context(task, source) — the narrative flow verb beside find — five tools, each still at most two parameters. Roster docs, argument checks and tools/list assertions updated.

All legs green: py 292/292, ts 225/225, java 329/329, js 285/285, cs 207/207, front_door 17/17, mcp ok.

[swapnilpaliwal-sd]

Merged apps/integration-0.1.9 into this branch (the conflicts were the surface trim vs context's return). Resolution: everything from the trim stands — find, graph, diff and install stay gone — and context stays surfaced as a first-class verb: the dispatcher's public help (which is the delegated list), the MCP roster (context, impact, path, tests), both SKILL.md copies, and the surfaces audit (fifth public verb, --source the one taught option; the off-surface refusal test loses context and gains a positive check).

After the merge: front_door 21/21, mcp ok, mcp_first 30/30, surfaces ok, python 302/302, java 333/333, csharp 208/208. typescript 25 and javascript 11 failing checks are the incoming const-object cases and fail identically on pristine apps/integration-0.1.9 — pre-existing on the base, not from this branch.