fills: nearest-tick rounding of raw bar prices (TradingView fill rule) - #163

… private fixtures (#375)

Squash of #375 (one commit on b802cee = v1.5.0; tree 18d3b6f). Its commit message follows verbatim.

tests: the ten test files that embed US-equity price data are private…

… rows (Form B)

Ten files leave the public tree for pineforge-private-fixtures (byte-exact
copies, commit 5abcf8a6): tests/test_ta_series_length_data.hpp,
test_ta_series_length_extremes.cpp, test_ta_series_length_supertrend.cpp,
test_ta_simple_length.cpp, test_ta_window_underapplied_data.hpp,
test_ta_window_underapplied.cpp, test_ta_window_round10_data.hpp,
test_ta_window_round10.cpp, test_tv_fill_rounding.cpp,
test_tv_fill_rounding_l4d.cpp. Their data headers and tables hold NYSE:F 15m
bars, NASDAQ:AAPL closes and TradingView indicator values on NYSE:F@15.

Six rows become private rows registered the #371 way (the real test when the
overlay supplies the source, a SKIP placeholder with the pinned literal
otherwise): test_ta_series_length_extremes, test_ta_series_length_supertrend,
test_ta_simple_length, test_ta_window_underapplied, test_ta_window_round10,
test_tv_fill_rounding_l4d. test_tv_fill_rounding (not the _l4d twin) was
already retired through L3A_LEGACY_OWNER_TEST_SOURCES and registers no row; its
source leaves and its name stays in that retirement list (REMOVE_ITEM of an
absent name is harmless).

The four source-bound rows are recorded in _pf_source_bound_private_rows so the
kernel profile, which has no source layer, stays correct without the files.
Pins: tests/private_fixture_rows.txt 368 -> 374 rows (and
PRIVATE_FIXTURE_ROW_COUNT), tests/private_fixture_paths.txt +10 F lines (10,206
files, 12,910 paths, 2,704 directories). Floors: release private count 300 ->
306 (absent floor 610 -> 604), kernel private count 28 -> 30 (only the
underapplied and round10 rows are source-free; absent floor 309 -> 307); full
totals 910 and 337 are unchanged.

Doc sentence reworded, named here for the doc-revert gate: "The full-run
release and kernel floors are the table's own numbers above (610 and 309 rows
absent, 910 and 337 with the overlay present); full runs do not exclude a
label." now says 604 and 307 for the absent floors; the two cells of the floor
table in docs/ci.md change the same way (610 -> 604, 309 -> 307).

None of our release-built assets contains these files: the prebuilt tarballs
are cmake --install with PINEFORGE_BUILD_TESTS=OFF plus LICENSE/NOTICE/VERSION
(release.yml, the build at line 399 and the pack step at 433); only GitHub's
automatic source archive carries tests/. Git history is not rewritten (owner
ruling 10-11). First added: test_tv_fill_rounding.cpp 7fc7347 (2026-08-25,
#163), test_tv_fill_rounding_l4d.cpp 73817c1 (2026-09-19, #254),
test_ta_window_underapplied* ade627e and test_ta_window_round10* c8993aa
(2026-09-06), test_ta_series_length_*, test_ta_simple_length.cpp 1e360bd
(2026-09-27).

Two existing gates carried the old numbers and move with the floors (found by
the Claude read of the first cut): scripts/test_ci_verify.py (the literal pins
(910,306)/(337,30) and absent floors 604/307, the kernel absent scenarios now
with 30 pinned rows, the explicit --min-tests 307 case, the one-row-ran case
337 - 29 = 308, the websocket allowance case 337 - 30 - 1 = 306; the shrunken
release registry 901/298 -> 603 still fails the 604 floor; dated measurements
stay as history) and tests/private_fixture_reads.evidence (the CMake line it
cites moved by -2: 3253 -> 3251, prose cites :3273 -> :3271, :3237 -> :3235,
:3249 -> :3247, :3279-3280 -> :3277-3278). A comment in tests/CMakeLists.txt
that cites the two gate suites follows (:2217/:2219 -> :2215/:2217).

Not run locally (no suites on the Mac); the two-world proof runs on lab remote.


Co-Authored-By: Claude Opus 5.5 (1M context) <[email protected]>