[NewMe505]

Summary

The bugs behind the graphcalc run's S5, where P4 halted twice on its environment and never on the math. Built by three parallel agents (one per fix, owned files); then an independent review of the whole diff, whose three findings are fixed here too.

Closes #69

Plan vs. done

Bug	Fix	Regression evidence (fails on main)
I-114: a Claude Code worker turn's python was SCIO's own environment, so P4 could not import the approved sympy	with_interpreter_first: the product interpreter's directory first on PATH (VIRTUAL_ENV set), for the worker's claude -p calls, every LocalSandbox.exec command and the local app start; the Windows bare-word lookup uses that PATH (review); BACKEND_NOTE says python is the product's	test_worker_turn_runs_with_the_product_interpreter_first_on_path, test_every_local_command_runs_with_the_product_interpreter_first_on_path, test_a_bare_command_word_resolves_through_the_product_first_path (Windows)
I-115: advisor cap hit (2/3, step S1a) when the per-step cap fired	the halt names the cap that fired	test_halt_reason_names_the_per_step_cap_when_the_total_has_headroom
I-116: skip dropped a feature for good	the owner's decision, skip defers: skipped stays open; FeatureRun.milestone, so a skip finishes a feature only in its own milestone; per-milestone cost counts its own runs; the release refuses while a feature is not done; the PR closes only passed features' issues; PRODUCT.md is re-read inside the progress commit (review) and at the release stop (review); older skips are stamped when a milestone closes (review)	test_a_skipped_feature_is_deferred_to_the_next_milestone_not_dropped (end to end), test_closing_a_milestone_stamps_runs_recorded_before_the_milestone_field, test_a_feature_added_at_the_release_stop_holds_the_release, and four more in the ledger entry
I-117: a subagent's edit through a shell script with \n	nothing damaged; guardrail 17 extended	—
I-118: the full check caught a shutil.which fake left at one argument after the review fix	the fake takes path; guardrail 38	the fixed test

Guardrails 35 to 37 (20 to 34 are PR #70's, so neither renumbers).

Test evidence

ruff check . && ruff format --check . && mypy && python -m pytest && python scripts/check_docs.py && python -m evals.runner --suite regression --provider fake --trials 5 --min-pass-pow-k 1.0
401 passed, 18 skipped in 879 s; docs-check: 0 problem(s); pass@1 1.00 pass^5 1.00 (10 tasks)   # on 8748b41, serial, clean worktree

CI on 8748b41 (run 40647941438): checks 413 passed, 6 skipped (the new Windows-only test skips on Linux); docker-live 15 of 15.
After merging main (the dashboard, #70; shared-file conflicts kept both sides): local check on f784381 500 passed, 18 skipped, pass^5 1.00; CI (run 41089330746) checks 512 passed, 6 skipped, docker-live 15 of 15.

Docs updated

• README (skip defers; status line)
• docs/ARCHITECTURE.md (the feature loop's skip)
• ADR (0011 amended: skip defers, the owner's decision)
• CHANGELOG [Unreleased]
• Docstrings for changed public APIs
• N/A: no user-facing or architectural change

Continuous improvement

• Issue ledger updated for every problem hit (ids: I-114 to I-118)
• Each bug fix was swept across the codebase (patterns in each entry)
• Milestone retrospective: not a milestone PR; the run's retrospective is docs: retrospective of the graphcalc live product run #63

Risk notes

• LocalSandbox.exec now gives every product command the product's interpreter first on PATH; SCIO's own tools stay after it.
• The worker tests stop at the process boundary: on the next live run, the coder's python -c "import sys; print(sys.executable)" should point under .scio/sandbox.
• Merging after feat(dashboard): a live, read-only dashboard: scio dashboard (#68) #70 touches the same shared files (CHANGELOG, ledger, LESSONS_LEARNED); guardrail numbers do not collide.

🤖 Generated with Claude Code