[eggmasonvalue]

What this changes

Renames. signal-sweep → idea-generation, sec-edgar-skill → sec-filings, market-scout → market-data, bottom-up-analyst → due-diligence. pitch-like-lou keeps its name. The cache signal-sweep-cache → idea-cache (env IDEA_CACHE_DIR). References are updated in package.json, bootstrap.mjs, the insider-scan workflow, pyproject, .gitignore, markdownlint config, AGENTS.md, context/, and the READMEs. In the skill lockfiles only the project-name line changed (edited by hand to avoid a uv format bump).

Where documentation lives (second commit; the first wrongly put gotchas in SKILL.md):

• SKILL.md: routes, commands, and pointers only.
• --help: flag semantics (filing-date selectors, 10-Q Part naming, 20-F items, ex-99.1).
• Error hints: recoverable mistakes (BRK.B → BRK-B).
• Report text: caveats (screen lag warning, transcript vendor note).
• references/: off-script detail. New files: sec-filings/references/{edgartools,ownership}.md, market-data/references/yfinance.md, idea-generation/references/hunting.md.
• A new CONVENTIONS.md rule records this.

due-diligence rewrite. Textbook material is cut. Six archetype files are merged into archetypes.md (edge, traps, valuation route, tell). The guides became normalization.md, valuation.md, and memo.md. guide_competitive.md is folded into SKILL.md. Ownership interpretation now lives only in sec-filings/references/ownership.md. Scope is unchanged: long-only, and banks, insurers, REITs, funds, and binary assets stay excluded.

pitch-like-lou. Tightened. The description now also triggers on a "stock pitch" or a "punchier rewrite".

Script fixes

• scan_insiders.py: the 13D scan returned nothing because EDGAR now uses SCHEDULE 13D/13D/A; querying SC 13D on 2026-09-15 returned 0 rows. The scan now queries both names, dedupes by accession (the index lists each schedule under both filer and subject), and takes the issuer from issuer_info (the old code could attribute a stake to the filer, e.g. FTK → ProFrac). Live test on 2026-09-15: 17 schedules.
• fetch_market_data.py: shows class shares and implied all-class shares separately when they differ. Prints a hyphen hint for dotted tickers.
• fetch_transcripts.py: adds the hyphen hint and a vendor-transcript note in each file.
• scan_market.py: reports for screens using lastclose52week* fields now carry a lag warning.
• fetch_filing.py, parse_financials.py, fetch_filings.py: --help text only.
• Removed list_headings.py (it was rg -n '^#').
• context/DECISIONS.md: the z-score entry now says 22 trading days vs the prior 252 rolling returns, not "30-day".

Known open issues (not fixed here)

• Yahoo's lastclose52weeklow/high.lasttwelvemonths screener fields lag the live quote. On 2026-10-31, a ≤ 1.0 query returned stocks 6–50% above their low. So near-52wk-low and fallen-from-grace give misleading results until they are redesigned. This PR only adds a warning.
• dcf.py and epv.py need Python 3.10+ (they crash on 3.9 outside the secpi venv).
• fetch_transcripts.py --latest trusts Yahoo's listing order. The Q&A split depends on an operator turn containing "question".
• scan_insiders.py clusters fall back to name matching when an owner CIK is missing.

Verification

• ruff check, ruff format --check, and markdownlint-cli2 pass.
• Live checks: SEC daily index (13D form names), edgartools company filings for both 13D spellings, yfinance 1.5.2 fields (units, BRK-B shares, ^TNX), screener lag, the BRK.B hint, and the screen-report warning.
• Not run end-to-end: fetch_filing.py/parse_financials.py after the --help edits (only --help was exercised), transcripts, theme search, and the conference scan.
• Luna (gpt-6-luna) eval, routing and dry-run plans only, no live calls. Old vs new descriptions on 20 prompts:
  • five prompts (EV/EBIT, 10y Treasury, options, consensus EPS, spin-offs) went from unrouted to routed;
  • the routing test revealed overlaps that were fixed (13F holders listed in both data skills);
  • dry-run plans caught doc/code mismatches that were fixed.
    These checks used the first-commit docs, before SKILL.md was slimmed, and were not re-run.