Repository navigation
Commit 39fc78c
Train: report-channel test bounds, stdin idle default, na string rule, capacity stop codes (#691)
Combines four open pull requests into one head on current main, so they merge and are verified once.
- #690 tests: assert behaviour instead of wall-clock bounds in the report-channel, deadline and compile-limit tests; the drain grace is observed as an event and the stall control's guard is an iteration count.
- #686 runner: the stdin feed's idle and message bounds default to none, with `tests/native_live_stdin_idle_e2e.py`.
- #688 na: read string-convertible types through the string rule (`include/pineforge/na.hpp`, `tests/test_na_generic_containers.cpp`).
- #684 report distinct-id and live-order capacity stops under separate failure codes.
Main (d6f03f2) plus the four heads, merged in that order: no conflicts, nothing edited by hand. Supersedes #690, #686, #688 and #684 (their branches are kept).
🤖 Generated with [Claude Code](https://claude.com/claude-code)
## The four pull requests in this train (bodies verbatim)
### #690 tests: assert behaviour instead of wall-clock bounds in the report-channel, deadline and entrypoint rows
Main is red on the post-merge Debug and arm jobs since the report-channel change: several tests asserted wall-clock upper bounds (for example "elapsed < 1.8 s"), which slower runners miss.
This change removes every elapsed-time bound in these tests, in both directions, and asserts behaviour instead: the terminal record's exit class, failure code, the reason (deadline or stall) and which grace expired.
- `tests/report_channel/test_contract.py`
- `tests/report_channel_deadlines/test_deadlines.py`
- `tests/entrypoint_limits/test_compile_limits.py`
The rows were checked red-first by breaking the end path: they fail when the supervisor does not end the run. Test-only; no product code changes.
🤖 Generated with [Claude Code](https://claude.com/claude-code)
### #686 runner: the stdin feed's idle and message bounds default to no limit
## What changes
The native runner's stdin feed (`pineforge-live run --feed -`) ended a run on its own after 15 seconds without data or 15 seconds to complete a line, and it counted its own processing time as "idle". Under the sanitizer builds one read of a burst took longer than the bound, so `native_live_ratio_interval_e2e` died with `native stdin idle or message timeout; 0 actions not sent` although data had just arrived.
1. **The clocks measure waiting, not processing.** After the pass that consumes the complete lines of a read, the idle clock restarts; the message-assembly clock restarts only if that pass consumed a message and a partial line remains. A continuation read that only appends bytes to an unfinished line does not refresh the assembly clock, so a sender that dribbles bytes cannot keep an unfinished line alive past a set message timeout (`runner/main.cpp`, stdin loop).
2. **Default change (user-visible): the stdin feed has no deadline unless you set one.** With neither `--feed-idle-timeout` nor `--feed-message-timeout` given, both default to 0 (no limit), each independently, so a quiet market is not an error. A set value still ends intake. For stdin the flags accept 0 (no limit) and any positive number of seconds (the 300 second upper bound is gone). Network feeds (`--feed-url`, HTTP and WebSocket) and local files keep the default of 15 and the 1 to 300 range, and refuse 0 with a message naming the flag. `--status-interval` is unchanged. A caller that relied on the 15 second default to end a run must close stdin or pass an explicit value (`CHANGELOG.md`, `runner/README.md`, `--help`).
3. **Tests.** `tests/native_live_stdin_idle_e2e.py` (new, registered in `runner/CMakeLists.txt`) covers: a burst whose reports exceed the pipe buffer while the test leaves stdout unread for two seconds (the runner blocks inside the consume pass for longer than the one second bound on any machine; red on main); a partial line and a valid line dribbled over several seconds with a one second message bound (must end intake) and with 0 (must commit); a two second silent gap with idle 0 (survives) and idle 1 (ends); the existing silence row; and a row with no timeout flags and a 17 second silence (must survive). The dribble rows start their clock only after the runner has committed a first bar, so they measure the bound and not the start-up speed of a sanitizer build. The test's own host-side waits are one parameter, `PINEFORGE_STDIN_TEST_HANG_GUARD_SECONDS`: unset, empty or `0` means no bound, a positive number of seconds is that bound, anything else stops the test naming it. `tests/native_live_service_e2e.py` replaces the old stdin refusal rows for 0 and 301 by rows that start and finish cleanly and adds exact-message refusal rows for HTTP and WebSocket feeds.
4. Two documentation anchors that cite `runner/main.cpp` line numbers follow the added lines (`docs/design/native-feature-parity.md`, `docs/live-ledger-budget.md`).
No kernel, strategy or report byte changes; only the runner's argument handling and stdin intake loop.
## Consumer sweep (who ran the stdin runner and relied on the 15 second default)
Read from the clones as they stand; "none found" is not "proved none".
- `pineforge-fund`, `pfund/feed/runner_proc.py:221-223`: passes `--feed -`, `--feed-idle-timeout 300` and `--feed-message-timeout 15` explicitly; unchanged by this PR. The feed process owns the runner's stdin.
- `pineforge-data`, `native/feed/README.md` (lines 66-67, 441-442, 547): documentation examples pipe the feeder into `pineforge-live run ... --feed -` with no timeout flags. Today a 15 second silence ended that pipeline; after this change it ends when the feeder exits and closes the pipe, or the process is stopped. Documentation only; no code in that repository launches the runner.
- This repository: no non-test caller of the stdin runner. Tests that start it on stdin pass explicit values (`tests/test_native_live_report_cli.cpp` and `tests/native_live_checkpoint_e2e.py`: 60 and 60; `tests/native_live_service_e2e.py`: 1/3 and 3/1), finish a finite input and close stdin (`tests/generate_native_live_opaque_report_golden.py`), or stop the child.
- No mention of `pineforge-live`, `--feed -` or the two flags in the application, compute, optimizer, MCP, lab, market, license, data-api, data-worker, dashboard and workflow repositories.
- Not readable from here, not guessed: deployed settings that live outside the repositories (job environments and entrypoints, container entrypoint variables), repositories without a local clone, and runtime configuration on the trading box.
## Evidence
Head tested: the tip of this branch (tree `bac743d2d258`). Its parent commit differs from it by one test file (the six host-side wait lines of `tests/native_live_stdin_idle_e2e.py`); the runner, CMake, README, CHANGELOG and every other test are byte-identical to that parent.
- Cloud verify (four steps: preflight, release, debug, sanitizers) on the parent commit: all four exited 0. Cloud preflight on the tip: exit 0 (60 stages, 0 advisory).
- Native-live on the tip: live-sanitizers 120 of 120 passed (`native_live_stdin_idle_e2e` 516.59 s, `native_live_ratio_interval_e2e` 179.05 s, `native_live_service_e2e` 138.94 s). Release native in four shards: shards 1, 2 and 4 passed (410, 371, 414 tests); shard 3 had exactly one failure, the known `test_native_live_report_memory` at its 160 s deadline, which passes alone three times (53.13, 53.43, 52.80 s) and does not touch the stdin runner. `native_live_stdin_idle_e2e` passed in shard 4 in 99.27 s, including the row with no timeout flags and a 17 s silence. live-tsan on the tip: exit 0, and `native_live_stdin_idle_e2e` alone under tsan three times: 213.62, 212.45, 213.50 s, each 1 of 1 passed. live-tsan on the parent commit: 120 of 120 (`native_live_stdin_idle_e2e` 291.84 s, `native_live_ratio_interval_e2e` 157.22 s, `native_live_service_e2e` 132.76 s).
- Red first: an earlier version of the new test, run against main's runner, failed at its first (burst) row with `native stdin idle or message timeout; 0 actions not sent`, twice, in 16.9 s; that version's dribble rows were then found to depend on start-up speed under sanitizers (they failed in live-tsan and live-sanitizers) and were rebuilt to start their clock after the runner has committed a first bar. The final test text was not re-run against main.
- Reviews (static reads): the first two commits, the review fixes, the default change and the test-host parameter each had an independent read; the last read (tip against its parent) found no P0, P1 or P2.
- Known class named, not hidden: shard 3's `test_native_live_report_memory` 160 s deadline red is the subject of a separate change that makes that test's budget a parameter.
## Seeing a stalled feed
With no idle or message limit set, a consumer detects a silent feed from the status file: `metrics.source_lag_ms` (runner/README.md:326) is refreshed every `--status-interval` even when no data arrives, so it grows while the feed is stalled. A dedicated `metrics.feed_idle_ms` (time since the last byte read) and the same no-limit default for the network feeds follow in a separate pull request.
🤖 Generated with [Claude Code](https://claude.com/claude-code)
### #688 na: read string-convertible types through the string rule, with rows for the double-conversion clause and the limit stated on the overload
## What
`pineforge::is_na` has a non-template overload for `std::string` (an empty string reads as na) and a generic template for everything else that is not arithmetic or convertible to `double`. A type that converts implicitly to `const std::string&` but is not a `std::string` (the shape of the engine's `SessionKey`) fell into the generic template and was always answered "present". This change reads such a type through the string overload instead:
- `include/pineforge/na.hpp`: the generic template gains the constraint `!std::is_convertible_v<const T&, const std::string&>`, so a string-convertible type is read through the conversion to `const std::string&` and the `std::string` rule ("is it empty") applies. `std::string_view` and engine string handles such as `ScriptString` are not `std::string`, never convert implicitly to it, and stay never-na; the comment on the overload now says so.
- The limit is stated on the overload itself: a plain value container (`std::vector<T, A>`, `std::map`, `std::unordered_map`) cannot hold an na reference, so `is_na` of one is always false, even for an empty container, and `std::vector` cannot tell an na array from an empty one. The comment on the unbound `ScriptArray` state also says what unbound means (no store behind it: default constructed or moved from).
## Tests
`tests/test_na_generic_containers.cpp` gains rows for:
- a type with only a non-const `operator double()`, called on a non-const lvalue (it must keep answering `isnan(double(x))`, which the existing `T&` clause of the template guarantees);
- a type that converts to `const std::string&`: an empty one is na, a non-empty one is not;
- `std::string_view{}`: never na.
## Files
- `include/pineforge/na.hpp` (header only: no struct, layout or ABI change)
- `tests/test_na_generic_containers.cpp`
Only calls of `is_na` on a type that converts to `const std::string&` without being a `std::string` change their answer. A search of `src/` and `include/` finds no call that passes the engine's `SessionKey` (the one such type in the engine, `include/pineforge/native_order_identity.hpp`) to `is_na`, so no engine behaviour, hash, state or report byte changes.
🤖 Generated with [Claude Code](https://claude.com/claude-code)
### #684 Report distinct-id and live-order capacity stops under separate resource labels
Both stops ended as scratch_capacity_exceeded with resource cohort, so a
consumer could not tell the 65,536-id historical tier from the 4,096-id
live cap without parsing the message. The historical count and byte
refusals now report cohort_ids_total and the live-order guard reports
cohort_orders_live. The slab region, the single-id length refusal and the
generic store limit keep cohort. No limit, message or refusal condition
changes. The two labels join the resource vocabulary in the failure
catalog, its registry and the generated documentation.
The memory-management failure table row for scratch_capacity_exceeded is
rewritten: the sentence "A slab region ran out, more than 4,096 cohorts are
live in the bounded profile, or an id is longer than 64 bytes in the bounded
profile" is replaced by one entry per resource label, and the sentence
"Admit larger maxima or shorter ids before another attempt" is reworded in
that row to also name closing or cancelling live orders sooner.
🤖 Generated with [Claude Code](https://claude.com/claude-code)
## Commits (verbatim, as git log prints them)
a296f35
Merge #684 into the train: report distinct-id and live-order capacity stops separately
Co-Authored-By: Claude Opus 5.5 <[email protected]>
f796b5e
Merge #688 into the train: na reads string-convertible types through the string rule
Co-Authored-By: Claude Opus 5.5 <[email protected]>
c99b140
Merge #686 into the train: runner stdin feed idle and message bounds default to none
Co-Authored-By: Claude Opus 5.5 <[email protected]>
1f3c5db
Merge #690 into the train: tests assert behaviour instead of wall-clock bounds
Co-Authored-By: Claude Opus 5.5 <[email protected]>
1f54ff8
tests: the drain marker is released by an event, not by a delay
The descendant now writes its marker once a file exists that the host
creates only after it has seen the producer gone while the descendant is
alive. No fixed wait remains in the fixture: a supervisor that stopped
draining at once never reads the marker, one that drains for its grace
does, decided by the order of events.
Co-Authored-By: Claude Sonnet 5.5 <[email protected]>
d1ee1b1
tests: the drain marker waits a scenario delay after the producer's exit
Written at once, the marker raced into the supervisor's last read even
when the finite grace returned immediately, so a grace of zero was not
caught. The descendant now writes it 250 ms after it sees the producer
gone, well inside the default grace: a grace that returns at once loses
it, one that drains for its grace reads it. The assertion stays an event
(the line is in the captured diagnostic), not a duration.
Co-Authored-By: Claude Sonnet 5.5 <[email protected]>
c24fc8c
tests: the drain rows accept the marker line in the no-corruption check
The captured diagnostic now holds two kinds of complete lines (the
descendant's progress line and the marker it writes only after the
producer exited); the exact-pattern comparison is replaced by a per-line
check that keeps the same guarantee: every complete line is one of the two
and a cut last line is a prefix of one.
Co-Authored-By: Claude Sonnet 5.5 <[email protected]>
b966b43
Merge main into the cohort resource label branch
Main now carries the parity-path and arena-bound script changes and the profile
environment anchor refresh. The merge is clean: the two changes share only two
documentation pages, and both sides' text is kept as it is. No source, test or
build file of the label change is touched by the merge.
Co-Authored-By: Claude Sonnet 5.5 <[email protected]>
5542f18
Merge main into na-string-convertible
Brings the current engine main (the parity adapter for a jailed host and the public wording in the design docs) into the branch that lets string-convertible types read through the na string overload.
Co-Authored-By: Claude Sonnet 5.5 <[email protected]>
6ce6bc4
tests: the drain grace is observed as an event, the control's guard is a count
The post-exit drain rows now assert that output the descendant wrote only
after the producer had exited reached the host through the supervisor, so
a grace that returns at once fails them without any duration. The report
channel mutation control's guard is a count of the host's own loop
iterations since it saw the stall (80 in the CTest row), not seconds;
unset means no guard and the row skips locally. CTest comment keeps its
line count, so the pinned CMake lines do not move.
Co-Authored-By: Claude Sonnet 5.5 <[email protected]>
0478add
Merge main into the stdin idle default branch
Brings the branch level with main after the parity path changes (#687).
No conflicts. Only docs/design/native-feature-parity.md changed on both
sides; the two edits are in different places and merged cleanly. Every
other file the merge brought in is untouched by this branch.
Co-Authored-By: Claude Sonnet 5.5 <[email protected]>
7d0994a
Merge main (the jailed parity adapter) into the no-wall-bounds branch
No conflicts. The only file both sides changed is tests/CMakeLists.txt:
main's hunk starts at line 7503, after every line this branch pins, so
the private fixture evidence pin at :3416 still holds.
Co-Authored-By: Claude Sonnet 5.5 <[email protected]>
e15d821
tests: re-pin the private fixture evidence row to the shifted CMake line
The report-channel rows added twelve lines to tests/CMakeLists.txt ahead
of the pooc close-fill view block, so the evidence token that
check_private_fixture_reads pins at :3404 now sits at :3416, and the line
cites in its reason text moved by the same twelve.
Co-Authored-By: Claude Sonnet 5.5 <[email protected]>
a8aaada
tests: the mutation control row runs every time with an explicit opt-in deadline
test_report_channel_mutation_control sets
PINEFORGE_REPORT_TEST_MUTATION_CONTROL_AFTER_STALL_SECONDS=4 itself. That
is not a limit on a real run: it is the test host's opt-in deadline
parameter given an explicit value (the product default stays no
deadline), and the row asserts the terminal record (HostWatchdogExpired),
not elapsed time. The row never skips, so ci_verify's skip policy is
untouched.
Co-Authored-By: Claude Sonnet 5.5 <[email protected]>
81e8da4
tests: the report-channel mutation control gets its own CTest row
A unittest-internal skip leaves the CTest row green, so the control's skip
was visible only inside one log. The control now runs as
test_report_channel_mutation_control, which CTest reports as Skipped when
PINEFORGE_REPORT_TEST_MUTATION_CONTROL_AFTER_STALL_SECONDS is unset (the
mutant never ends without a deadline); the contract row deselects it.
Co-Authored-By: Claude Sonnet 5.5 <[email protected]>
1990737
tests: no elapsed assertion in either direction; the mutation control's deadline is a parameter
The deadline suite drops every lower bound on elapsed time and on the
fixtures' own tick gaps (and the re-arm mutation half that depended on
one), and the timing data nothing reads any more. The terminal record,
the moved bytes and the syscall observations stay the assertions. The
report-channel mutation control takes its deadline from
PINEFORGE_REPORT_TEST_MUTATION_CONTROL_AFTER_STALL_SECONDS; unset means
no deadline, and the row then reports itself skipped because its mutant
never ends by design.
Co-Authored-By: Claude Sonnet 5.5 <[email protected]>
f93075b
tests: pin the doctored-grace controls to the printed limits record, drop a dead assertion
The two doctored-default controls now match the limits dict the assertion
prints ('effective': 3000) instead of any 3000 in the failure context. The
assertion after the ready loop in the entrypoint limits row can no longer
fail and is removed.
Co-Authored-By: Claude Sonnet 5.5 <[email protected]>
ab42036
tests: the last fixed wall limits in the same directories go
The class sweep found wall limits the earlier commits did not touch: the producer fixture gave up
after 15 s waiting for the host to observe a stopped validator; the entrypoint blocked-stderr row bounded
its ready loop by the compile wall; test_off_identity.py carried a 600 s watchdog (and its fd-sink and
wait deadlines, an exception wrapper and a 5 s wait after kill) and a 30 s git timeout; the host had a
30 s git timeout, a 10 s timeout on the reference run, a 1 s join of its exit-observer thread and a
2 s wait after kill. They now wait without a limit, as the rest of the three directories do (a real hang
is caught only by the platform's own maximum job time).
Not touched, on purpose: the scenario durations of the fixtures (a descendant that lives N ms, a CPU
spinner for N ms, a diagnostic reader that starts after N ms) and the host's optional guard whose
default is none: they are inputs of the scenarios, not limits on a run.
Co-Authored-By: Claude Sonnet 5.5 <[email protected]>
0ad3b43
tests: the doctored-default controls pin their failure; the unused bound_ms parameters go
The two doctored-default controls (post-exit drain and stop grace) now expect an AssertionError whose
text names the doctored value (3000), so only the limits-record mismatch passes them, not any
unrelated assertion or setup failure.
The deadline helpers single, write_case, read_case and relay_case no longer take a bound_ms parameter
(it had no use left once the external waits were removed), and every call site drops the argument; an
unused bound parameter would read as a bound that still exists.
Co-Authored-By: Claude Sonnet 5.5 <[email protected]>
cca6659
tests: the finalizer observer builds on arm64 and says why when it does not; dead waits and a false comment go
The C observer that the report-channel rows load into the validator used the raw pause syscall
(SYS_pause), which aarch64 does not have, so `cc -Werror` failed on the arm64 runners, and the build
helper swallowed cc's stderr, so the failure showed only as a CalledProcessError. The observer now
uses ppoll with no descriptors and no timeout where SYS_pause is absent (it blocks until a signal,
as pause does; x86_64 keeps SYS_pause), and the helper raises an AssertionError that carries the cc
command and its stderr. The helper's 30 s build timeout is removed too.
Also: the cleanup that reaps a fixture child no longer fails after two seconds (it waits as the
earlier reap does); Process.line() and Process.result() in the deadline harness lose their timeout
parameter and the "external wall bound" branch (no caller passes one), and the unused start time
goes; and the comment that named a "no-new-output guard" now names what the repository has, the
platform's own maximum job time (the workflows set no timeout-minutes).
Co-Authored-By: Claude Sonnet 5.5 <[email protected]>
c4a4af6
tests: the two doctored-grace controls now reject the doctored default through the limits record
test_post_exit_drain_default_assertions_reject_a_doctored_grace and
test_stop_grace_default_assertions_reject_a_doctored_grace expected the removed
"returned too late" timing messages. A supervisor whose default grace is doctored from 1000 ms to
3000 ms is still rejected, by the row's own behaviour assertion: the terminal record's limits
object shows the effective default (1000 expected, 3000 seen). They now expect an AssertionError
from the doctored run, and no longer name a timing message.
Co-Authored-By: Claude Sonnet 5.5 <[email protected]>
56a8ac6
tests: the report-channel, deadline and entrypoint rows assert behaviour, not wall time
No test bounds elapsed time any more. Two rows failed on GitHub's Debug jobs for exactly that
reason (test_finite_stall_expires_through_supervisor_diagnostic_pipe: 1.849 not less than 1.8;
test_paused_validator_is_ended_by_supervisor_wall: 3.187 not less than 3.15). Removed, in both
directions, from tests/report_channel/test_contract.py (cancel delay and exit-after-cancel,
finite stall, wall exit, diagnostic stall, validator wall, elapsed_seconds limits, drain grace,
producer and validator stop grace, process-exit polling allowance, the hold delay, the
delayed-diagnostic floor and the one-second fixture-child reap wait), from
tests/entrypoint_limits/test_compile_limits.py (the elapsed windows around the compile wall, the
outside timeouts of 4, 5, 7 and 30 s and the 180 s hang-guard constant) and, for the upper
bounds, from tests/report_channel_deadlines/test_deadlines.py (assert_timing's upper window and
every external wait of the harness: result(N), line(N), wait(timeout=2), the pause wait, the
EAGAIN loop bound and the fixture build timeout). The harness now waits until the fixture ends;
a real hang is caught by the CI runner's own no-new-output guard.
Kept: every behaviour assertion (terminal record, exit class, failure code, reason deadline or
stall, which grace expired, stage, limits requested/granted/effective, process tree gone, bytes)
and pure order comparisons of the host's own timestamps (cancelled_at before exited_at is
causality, not a bound). Deleted, not emptied: check_timing and the old assert_timing (the
deadline suite keeps assert_not_early, a deadline must not fire before its time, and the
re-arm row keeps its one explicit lower assertion). The mutation control keeps its own finite
guard counted from the stall: HostWatchdogExpired firing is the feature under test there.
Co-Authored-By: Claude Sonnet 5.5 <[email protected]>
94e841e
Merge main into na-string-convertible
Brings the current engine main into the branch that lets string-convertible types read through the na string overload.
The merge has no conflict; the branch's two files (include/pineforge/na.hpp and tests/test_na_generic_containers.cpp)
are unchanged by it.
Co-Authored-By: Claude Sonnet 5.5 <[email protected]>
4ea1174
Merge main 59050b4 into the branch
One conflict, in CHANGELOG.md: both sides added entries at the head of the
Unreleased section. The resolution keeps every line of main's entries unchanged
and keeps this change's default-change entry; no other file conflicted.
Co-Authored-By: Claude Sonnet 5.5 <[email protected]>
41482b3
Merge main into the cohort resource label branch
Main now carries the opt-in supervised report channel, the compile-step limits
and the deadline helpers for the container runner. The merge is clean: the two
changes share only two documentation pages, and both sides' text is kept as it
is. No source, test or build file of the label change is touched by the merge.
Co-Authored-By: Claude Sonnet 5.5 <[email protected]>
33d620f
Merge main into na-string-convertible
Brings the current engine main (memory caches, equity-curve window, cohort exposure fold, admission batches) into the
branch that lets string-convertible types read through the na string overload. The merge has no conflict; the branch's
two files (include/pineforge/na.hpp and tests/test_na_generic_containers.cpp) are unchanged by it.
Co-Authored-By: Claude Sonnet 5.5 <[email protected]>
9426ac8
Keep the two symbol-claimed citations of the memory bound table when naming the id table labels
The memory bound table's cohort row cited the cohort record that holds the five
summary words, and the entry-rank row cited the entry-rank key table member
with the pair of calls that insert into it. The merge of main 7944e10 replaced
both with single-line citations that name no symbol, so this commit is a stated
revert of 7944e10 for those two citations: it puts main's wording back, the
record at its struct and the member at its require-insert and insert pair, next
to the link record and end offset citations the label change added. Only these
two table rows change.
Co-Authored-By: Claude Sonnet 5.5 <[email protected]>
e197275
tests: the stdin idle test's host-side wait bounds are one named parameter, default none
native_live_stdin_idle_e2e had numeric host-side bounds (select waits of 180 s and
communicate bounds of 180, 60 and 15 s). A slow sanitizer build must not be failed
by them and they must not be raised to make a slow box pass (no time limit
anywhere, tests included). They are now ONE parameter,
PINEFORGE_STDIN_TEST_HANG_GUARD_SECONDS: unset, empty or 0 means no bound; a
positive decimal number of seconds is that bound; anything else stops the test at
start naming the parameter. All six sites (three select waits, three communicate
calls) use it. The runner's own deadlines under test stay as they were: the
silence row's wait of 15 s, the poll() checks and the named stimulus sleeps.
Co-Authored-By: Claude Sonnet 5.5 <[email protected]>
3ab8e63
docs: anchors into runner/main.cpp follow the four lines the stdin default added
The stdin default block in runner/main.cpp moves the calc_on_every_tick refusal
from line 307 to 311 and ledger.checkpoint_recovered_report from 1116 to 1120;
docs/design/native-feature-parity.md:124 and docs/live-ledger-budget.md:146 cite
the new lines. (runner/README.md:220 cites 1053 for the same call and was
already stale on main; it is not touched here.)
Co-Authored-By: Claude Sonnet 5.5 <[email protected]>
f9a77d1
runner: the stdin feed's idle and message bounds default to no limit; only a set bound ends intake
With neither --feed-idle-timeout nor --feed-message-timeout given, the stdin
feed now has no deadline (0), so a quiet market no longer ends a run after 15
seconds. A set value behaves as before; the network feeds (HTTP, WebSocket) and
local files keep the default of 15 and the 1 to 300 range and still refuse 0.
The help text and runner/README.md say so, and CHANGELOG.md states the default
change plainly as user-visible.
tests/native_live_stdin_idle_e2e.py gains a row with no timeout flags: after one
committed bar and a 17 second silent gap (longer than the old 15 second default)
the runner is still alive, accepts a second bar and exits 0 with two inputs.
Co-Authored-By: Claude Sonnet 5.5 <[email protected]>
c1129b4
tests: the stdin dribble rows start their clock only after the runner has proved it is ready
The dribble rows of native_live_stdin_idle_e2e started their stimulus clock
right after launching the runner. A sanitizer build needs seconds before its
loop reads stdin; it then read the whole dribble in one read, its message
clock started at that late read, and the check at the end of the 3.2 s dribble
found it alive (partial-dribble failed in live-tsan and live-sanitizers while
the Release build passed). The helper now writes one valid bar and reads its
committed report first, exactly as the gap rows do, and only then dribbles; the
dribbled line is the next bar, so a timeout row ends with one committed input
and the unlimited row with two. The wait after the dribble is a bound of 60 s
instead of 15 s, not an added wait. Nothing else in the rows changes.
Co-Authored-By: Claude Sonnet 5.5 <[email protected]>
f762657
docs: live-ledger-budget anchors the post-replay checkpoint_recovered_report call at its new line
The stdin timeout validation block added to runner/main.cpp in this change moves
ledger.checkpoint_recovered_report(recorded) from line 1108 to line 1116.
docs/live-ledger-budget.md:146 cited runner/main.cpp:1108; it now cites 1116.
Found by the review of the delta, not by the doc-anchors stage.
Co-Authored-By: Claude Sonnet 5.5 <[email protected]>
7cb92a8
Re-pin the memory bound table's id table citations after the main merge
Main's id table restructuring moved the end-offset member and the cohort and
entry-rank tables. The two cohort rows now cite the end offset at its new line,
the cohort table member and the entry-rank membership check at their new lines.
Only the numbers change.
Co-Authored-By: Claude Sonnet 5.5 <[email protected]>
c91a744
docs: native-feature-parity anchors the runner's calc_on_every_tick refusal at its new line
The stdin timeout validation block added to runner/main.cpp in this change moves
the calc_on_every_tick refusal from line 299 to line 307. The doc-anchors stage
of preflight reported exactly this one bad anchor on the previous head
(docs/design/native-feature-parity.md:124, main.cpp:299 -> main.cpp:307).
Co-Authored-By: Claude Sonnet 5.5 <[email protected]>
2404b5c
Merge commit '69cddc9d102d2ccf5605231cc508b82c08c7081c' into [internal]
c4a8cc4
runner: a continuation read must not refresh the stdin message-assembly clock; the stdin idle test is deterministic and covers the dribble cases
Review of the two earlier [internal] commits found that the end-of-pass restart
refreshed message_started after every read, including a read that only appended
bytes to an already partial line. A sender dribbling one byte per interval
without a newline therefore kept the assembly clock young for ever (the 1 MiB
frame limit was the only stop). The restart now happens only when the pass
consumed a message in this read; last_data is still restarted unconditionally,
so processing time never counts as idle and a truly silent pipe still times out.
tests/native_live_stdin_idle_e2e.py: the burst row no longer depends on how fast
a machine is. The test does not read the runner's stdout for two seconds after
the burst (a slow stdout consumer, the production cause), so the runner blocks
inside the consume pass for longer than the one second bound on any machine,
and the test asserts that more than 262144 report bytes were produced before EOF
so the construction cannot pass vacuously. New rows: a partial line dribbled one
byte per 0.4 s must time out on the message bound; a valid line dribbled over
three seconds times out with a message bound of 1 and commits with a bound of 0;
a two second silent gap passes with idle 0 and times out with idle 1; the
existing silence row is kept. runner/README.md says when a partial line must
finish.
Co-Authored-By: Claude Sonnet 5.5 <[email protected]>
7944e10
Merge main into the cohort resource label branch
Main now carries the cohort exposure fold, the id table restructuring and the
re-pinned documentation anchors. The one conflict was the memory-bound table in
docs/design/arena-bound.md: the two cohort rows keep the labels that name the
link record and the end offset, the other rows take main's side, and every
citation number in the table is re-pinned in a follow-up commit from the
checker's list.
Co-Authored-By: Claude Sonnet 5.5 <[email protected]>
8232ae8
runner: --feed-idle-timeout and --feed-message-timeout accept 0 (no limit) and any larger value on the stdin feed
The two flags were parsed as 1..300 seconds for every feed, so the stdin bounds
could not be switched off. They are now parsed as unsigned seconds; for the
stdin feed 0 means no limit and there is no upper cap. The network feeds (HTTP,
WebSocket) and local files keep 1..300 and refuse anything else after the feed
kind is known, with a message that names the flag and says the bound stays for
network feeds. The elapsed time is converted to whole seconds before the
comparison, so a large accepted bound cannot narrow. --status-interval keeps
1..300 and the default stays 15 seconds.
tests/native_live_service_e2e.py replaces the stdin 0 and 301 refusal rows by
rows that start and finish cleanly (0, 301 and the largest unsigned value, a
padded partial line included) and adds exact-message refusal rows for HTTP and
WebSocket feeds; the malformed, status-interval and local-file refusals stay.
The help text and runner/README.md describe the new meaning.
Co-Authored-By: Claude Sonnet 5.5 <[email protected]>
31c6129
runner: the stdin feed's idle and assembly clocks no longer count the engine's own processing time
The stdin loop of the native runner stamped last_data when bytes were read and
then consumed every complete message of that read before checking the clock
again. When one pass took longer than --feed-idle-timeout (15 s by default; the
sanitizer builds of native_live_ratio_interval_e2e did), the next loop head threw
"native stdin idle or message timeout" although data had just arrived. After the
consume pass the loop now restarts last_data, and the message clock for any
partial line left over. A feed that really is silent times out exactly as before.
tests/native_live_stdin_idle_e2e.py is registered next to
native_live_ratio_interval_e2e. One row pipes a burst whose processing exceeds
the one second bounds and requires exit 0, every input committed and the final
report equal to the independent batch oracle; the other keeps stdin open and
silent after one bar and requires the exact timeout diagnostic.
Co-Authored-By: Claude Sonnet 5.5 <[email protected]>
f058d4e
Merge engine main (the output series limit parameter) into the string-convertible na branch
Brings the branch up to date with main so the strict merge ruleset holds. The branch still reads
string-convertible types through the string overload of na and adds the rows for the reference
double-conversion clause and for string-like and string_view inputs; this merge changes none of those files.
Co-Authored-By: Claude Sonnet 5.5 <[email protected]>
36439bf
Name the symbols of the two id table citations in the arena bound note
The two table rows cite a line range of the id table header without naming
the symbol the range shows, which the anchor check refuses. Name the node
layout and the end offset the ranges point at.
Co-Authored-By: Claude Sonnet 5.5 <[email protected]>
9e94770
Merge main into the capacity label branch
The failure catalog changed on main since this branch's base (the output
series limit fields), so the generated registry and the catalog diff record
are re-derived on top of main's catalog with the two new capacity labels
(registry offsets after the resource vocabulary shift by two; the diff
record keeps main's release origin and records both changes). The
documentation anchors that cite the catalog and the id table are re-pinned.
Co-Authored-By: Claude Sonnet 5.5 <[email protected]>
891ecdd
Merge engine main (scratch-directory wording) into the string-convertible na branch
Brings the branch up to date with main so the strict merge ruleset holds. The branch still
reads string-convertible types through the string overload of na and adds the rows for the
reference double-conversion clause and for string-like and string_view inputs; this merge
changes none of those files.
Co-Authored-By: Claude Sonnet 5.5 <[email protected]>
53fda49
na: read string-convertible types through the string rule; rows for the double-conversion clause; state the container limit on the overload
A type that converts implicitly to const std::string& (such as a session key)
used to reach the generic template and answer "present" even when its string
was empty. The template is now constrained away from such types, so they take
the string overload and are read as strings. String views and other string
handles that are not std::string are documented as never na.
The constraint that keeps a class with a non-const conversion to double on the
old is_na(double) path had no row: a class with only a non-const operator
double, called on a non-const lvalue, would have answered false instead of the
isnan result. Add rows for a NaN and a non-NaN value through a non-const lvalue,
plus rows for the string-convertible type and for a string view.
The limit that a plain std::vector, std::map or std::unordered_map is never na
(an empty one included) is now stated on the generic overload itself, and the
unbound ScriptArray state is described as having no store behind it.
Co-Authored-By: Claude Sonnet 5.5 <[email protected]>
4f88b15
Report distinct-id and live-order capacity stops under separate resource labels
Both stops ended as scratch_capacity_exceeded with resource cohort, so a
consumer could not tell the 65,536-id historical tier from the 4,096-id
live cap without parsing the message. The historical count and byte
refusals now report cohort_ids_total and the live-order guard reports
cohort_orders_live. The slab region, the single-id length refusal and the
generic store limit keep cohort. No limit, message or refusal condition
changes. The two labels join the resource vocabulary in the failure
catalog, its registry and the generated documentation.
The memory-management failure table row for scratch_capacity_exceeded is
rewritten: the sentence "A slab region ran out, more than 4,096 cohorts are
live in the bounded profile, or an id is longer than 64 bytes in the bounded
profile" is replaced by one entry per resource label, and the sentence
"Admit larger maxima or shorter ids before another attempt" is reworded in
that row to also name closing or cancelling live orders sooner.
Co-Authored-By: Claude Sonnet 5.5 <[email protected]>
## Merge note
Squash of the T7 train (#691), which supersedes #690, #686, #688 and #684 (their branches are kept). Gates on head a296f35: pineforge/parity = sweep exp-engine-pr691-a296f352-20261101, 134/134 cases, neutral PASS (7,989 comparable, coverage lost 0, hard regressions 0, leavers 0; verdict 1e19e178); pineforge/verify = full 8 rj-20261101t064927-314e44 on tree 19f53d5; native-live on a296f35 = live-sanitizers and live-tsan exit 0, native shards 1, 2 and 4 exit 0, shard 3 one known timing-class failure (test_native_live_report_memory) that passed solo three times on the same box. GitHub checks on the head: green. Internal tracking ids in two commit messages are replaced by [internal]; nothing else in them is changed.
Co-Authored-By: Claude Opus 5.5 <[email protected]>1 parent d6f03f2 commit 39fc78c
36 files changed
Lines changed: 885 additions & 515 deletions
File tree
- ci
- docker
- docs
- design
- pages
- include/pineforge
- detail
- runner
- src
- tests
- entrypoint_limits
- report_channel_deadlines
- report_channel
| Original file line number | Diff line number | Diff line change | |
|---|---|---|---|
| |||
7 | 7 | | |
8 | 8 | | |
9 | 9 | | |
| 10 | + | |
| 11 | + | |
| 12 | + | |
| 13 | + | |
| 14 | + | |
| 15 | + | |
| 16 | + | |
| 17 | + | |
| 18 | + | |
| 19 | + | |
| 20 | + | |
| 21 | + | |
| 22 | + | |
| 23 | + | |
| 24 | + | |
| 25 | + | |
10 | 26 | | |
11 | 27 | | |
12 | 28 | | |
| |||
| Original file line number | Diff line number | Diff line change | |
|---|---|---|---|
| |||
1264 | 1264 | | |
1265 | 1265 | | |
1266 | 1266 | | |
1267 | | - | |
1268 | | - | |
1269 | | - | |
| 1267 | + | |
| 1268 | + | |
| 1269 | + | |
1270 | 1270 | | |
1271 | 1271 | | |
1272 | 1272 | | |
| |||
1533 | 1533 | | |
1534 | 1534 | | |
1535 | 1535 | | |
1536 | | - | |
1537 | | - | |
1538 | | - | |
1539 | | - | |
1540 | | - | |
1541 | | - | |
| 1536 | + | |
| 1537 | + | |
| 1538 | + | |
| 1539 | + | |
| 1540 | + | |
| 1541 | + | |
1542 | 1542 | | |
1543 | 1543 | | |
1544 | 1544 | | |
| |||
1573 | 1573 | | |
1574 | 1574 | | |
1575 | 1575 | | |
1576 | | - | |
| 1576 | + | |
1577 | 1577 | | |
1578 | 1578 | | |
1579 | 1579 | | |
| |||
0 commit comments