tests: recorded US-listed, futures, forex and metal tapes move to the private fixtures (LVB-60)

[luisleo526]

What

Recorded market data of US-listed stocks, US futures, forex and spot metals no longer lives in this repository. 85 files leave the tree:

• 24 files in seven directories (US-listed stock tapes under tests/fixtures/{session_ismarket,host_member_names,lazy_held_source}/),
• 60 files in fifteen directories (futures, forex and spot-metal tapes under tests/fixtures/session_ismarket/ and tests/fixtures/c7_tv_evidence/tv-eurusd),
• one trade list under tests/fixtures/tail_f_tv/.

They move byte for byte to the private fixtures repository, and the lab overlays them onto the paired engine tree under tests/codegen-fixtures/.
The tests that read them go through one helper, tests/_private_fixtures.py::private_fixture(rel):

• the file is there: the test runs exactly the assertions it ran before;
• the file is not there and no overlay was asked for: the test skips with private fixtures not present: <path>;
• the file is not there although PINEFORGE_PRIVATE_FIXTURES is set: AssertionError, never a skip, so a broken overlay cannot turn a private test into a silent pass.

Prose that quoted dates, counts, hashes or tickers of the moved tapes (READMEs, AGENTS.md / CLAUDE.md, comments and docstrings in 17 files) now says "a US-listed stock" or describes the outcome in words. Comment and docstring edits are AST-equal to before (checked per file).

How the tests changed (no assertion lost)

• Seven test modules read tapes through the helper (test_e2e_session_ismarket, test_e2e_session_windows, test_e2e_session_history, test_e2e_tail_f_tapes, test_host_member_names, test_e2e_lazy_held_source and test_e2e_cgint4_compositions), plus three recursive source scanners (test_diagnostic_codes, test_input_metadata_receipt, test_e2e_security_epoch_int64) whose private sources become one visibly gated case each.
• An explicit tuple of private slugs replaces substring routing, so a new private slug can never be read from the public folder by accident.
• Assertions that can only hold with every tape present moved, with their original text, into private-gated tests: the set totals ({"hm-g236": 1138, "cgim-flags": 1138}), the daily chart's whole-session flags, the nine-tape first/last map with its break/reopen loop. The public export tests assert each remaining public tape on its own. Public test ids do not change.
• The helper is specified first (a red run is on record: the helper test fails at import on the commit before the helper exists) and its tests fail, not skip, on an unexpected skip.
• The per-source host-member test asserts exactly one source transpiled, so a private source that stops compiling cannot pass silently.

Verification

Two worlds on the lab, paired with engine main (companion tree 7da487ab):

world	result
absent (private files not overlaid)	pytest 9781 passed, 221 skipped, 0 failed (10,002 tests collected); corpus compile 312/312; gate passed
overlay (private fixtures overlaid)	pytest 9936 passed, 66 skipped, 0 failed (10,002 tests collected, the same count as the absent run); no skip names a missing private fixture

Both rows are for the head of this PR. The overlay job first ran on a build host that a foreign process had loaded to 64 cores and timed out; it was resubmitted and ran on another host. That host has no clang++, so the 24 host-member cases (22 private sources and 2 public) skip with "needs clang's AST dump" in this run. They ran, and passed, in the overlay run of the previous head of this branch (23cd0017: pytest 9962 passed, 36 skipped, 0 failed, 9,998 tests collected, no private-fixture skip, on a host with clang++); tests/test_host_member_names.py and the private sources did not change between the two heads. The other 42 skips of the overlay run are environment skips of the kinds the absent run also has (the pre-lane codegen is not in the checkout's history, empty parameter sets, string-settings population sources). The difference between the two worlds is the tests that need the moved tapes: they skip in the absent world and run in the overlay world.
Two independent Claude reads of the diff (read-only, raw permission records empty) found no faked pass, no lost or loosened assertion and no private read at import or collection; their fixes are in.

Not in this PR

Git history is not rewritten. ta.py:764 and NO_HISTORY_READ_REASON keep a dated string (emitted text; later, with a regeneration note). Wording of the exporter tool name in a few pre-existing lines is a separate change.

🤖 Generated with Claude Code