[khushalsonawat]

Review only commits after 73e9170. This PR carries #276 (TH-9490, 73e9170), which isn't in feat/TH-8804-source-subfolder yet.

Summary

When the contract is saved, the kit follows the agent's imports to the code it runs, so the browser check and instrumentation see shared code outside the agent's folder. That walk skipped some code without a word, and ran out of reading on ordinary agents.

• The walk says when it leaves code unread (stopped, which the save logs as a WARNING and bundling records as truncated in credentials.json). This covers:
  • a .py holding import whose decoded reading isn't parsed: too big, past the total, or newer syntax;
  • a code file over the 1,000,000-byte read limit, or a link to one;
  • a folder link in the agent's folder that leads inside the repository;
  • an import name dropped for its length or depth, scripts included;
  • a link target past 128 parts;
  • a file in a codec decoded first, left undecoded once the walk's decode total ran out.
• Nothing else changes stopped: a file with no import, a standard-library name, a file Python itself refuses, a data file a script imports, and a folder link out of the repository.
• The save's walk gets its own 512 KB parse total. It was the 256 KB the other scans share, and those keep 256 KB.
• Each reading of a file read two ways is charged as it's parsed. This replaces fix(harness): decode a Python file before searching it when its codec can spell a name (TH-9490) #276's charge-once rule, which let a crafted pair cost twice the total. The parse total never refuses a file the old 256 KB let through.

Why: an import Python runs could hide behind a file the walk couldn't parse, and every ordinary subfolder agent over 256 KB of Python was cut on every save.

TH-9543 lands before TH-9512, which re-points its walk-stop test when it merges this.

Known limits:

• The four real subfolder runs measured stay cut: futureagi's ai_tools and agentic_eval, and the kit's src/fi/alk/harness and examples/frameworks. They reach 4-32 MB of Python and more than 5,000 import names. Reading them fully is TH-9445's (one capped subprocess and one deadline for the save).
• At the save, any reached .py over 128 KB that holds an import now says the walk stopped: 53 tracked files in the futureagi backend, 18 in the kit's src/.
• Honest agents up to about 500 KB of reached Python aren't stopped: measured, 305 and 401 KB are not; 517 KB is.
• The walk reads a decoded file only from its bytes, the way Python imports it, not also as text, the way it runs a script. A file in a codec outside the raw-safe set whose decoded text holds a CR can read differently the two ways, so an import only the script reading sees isn't followed. TH-9520 reads both ways for the browser scan, instrumentation and tool evidence; the walk doesn't yet.

Pre-existing, for Khushal's read (TH-9445):

• On 3.11, the combined worst crafted layout is over the 4 s line before this change: 6.4 s on feat/TH-8804-source-subfolder, 6.3-8.0 s with fix(harness): decode a Python file before searching it when its codec can spell a name (TH-9490) #276, and 7.9-8.3 s with this change. On 3.12 it's 3.6-4.4 s before, and 4.2-4.8 s with this change, under the 5 s line Khushal set for feat(harness): build an agent from a repo's subfolder #225.
• The walk alone with its other limits spent (1,500-4,500 missing names through a link chain, plus files only the walk parses): 1.7-2.6 s on 3.12 (base 1.1-1.9 s), and 3.4-3.9 s on 3.11 (base 1.8-2.1 s; 3.75-3.90 s on base's unicode-escape shape). The Coordinator sent this to TH-9445 under the same 5 s line: an extension of Khushal's ruling.
• Not measured on 3.13, the hosted image's Python: no environment here has 3.13 with the kit's dependencies, and installing one waits for Khushal.

Verification

• uv run ruff check .: not run on the whole repo.
• ruff check on the changed files passes. Focused pytest target, the seven changed-area files: 1532 passed on 3.12.14 and 3.12.3, and 1506 passed plus 26 skipped on 3.11 (25 of fix(harness): decode a Python file before searching it when its codec can spell a name (TH-9490) #276's are 3.12-only).
  • tests/harness/test_agent_type_refusal.py: one test per case above, plus a test that pins 512 KB exactly.
  • tests/test_harness_credentials.py: a walk that can't parse a file marks the manifest truncated.
  • tests/harness/test_python_source_codecs.py: fix(harness): decode a Python file before searching it when its codec can spell a name (TH-9490) #276's charge test now charges each reading.
  • Run with pytest -q on each of those files.
• sdk-smoke.yml's list on 3.12.14: 3689 passed, 3 skipped, 2 deselected. The named list on 3.11: 4179 passed, 32 skipped and 16 failed, the same 16 that fail on feat/TH-8804-source-subfolder, none about this walk.
• TypeScript build/test: not touched.
• agent-learn release-proof: not run.

Checklist

• Public API names stay under agent_learning.*, agent-learn, or @future-agi/agent-learning-kit
• New user-facing workflow has docs or an example: no user-facing workflow
• New release claim has executable gate coverage
• No hosted-service dependency was added to release-gated examples
• Security-sensitive output is redacted or justified

[cdileep23]

Please check the comments.