v0.8.1 streaming API: verified loose ends before release

[mario4tier]

Review of the v0.8.1 streaming API for consistency, usefulness and test coverage, across C/Rust/Java/C#. Baseline 2ef8a8cfd (post-#384).

80 candidate findings were raised; the 16 rated blocker were then adversarially verified against the shipped libta-lib.so.0.8.1. None survived as a blocker — 1 refuted, 11 should-fix, 4 nice-to-have. The 64 remaining are unverified and listed last; given that verification downgraded 16 of 16, treat them as leads.

Framing: only C has ever shipped, and the C streaming tier is new here — so its shape is still free, and this is the last moment it is.

A. API shape — settle before the streaming tier freezes

• Composed Update is not atomic. Ruled: documented, not fixed — PR docs(errors): intermediate overflow is undefined, said once (#386) #403. The mechanism is real and all five sites still tear at 7518b3343, measured with a control-paired probe (two handles primed identically, A fed one extra bar that Update rejects, then the same tail to both): KC returns TA_SUCCESS with wrong values for 7,019 bars before re-converging; MACDEXT and STOCHRSI reject 20,000 of 20,000 later bars and never recover; STOCH and STOCHF commit their own ring/extrema/today before the first sub and tear the window. PVO/PPO/APO/BBANDS/STDDEV/KDJ/ADXR are clean — their subs get the raw bar, already vetted by the parent's U3 (KDJ inherits STOCH's tear transitively).
  Two claims above did not survive the re-measurement. (1) "Fix STOCH/STOCHF and KAMA divide by a value their guard never tested: TA_SUCCESS with inf/nan output for well-formed OHLC #390 and 4 of the 6 sites become unreachable" — no: fix(stoch,kama): divide by the value the guard tests, and bound the efficiency ratio (#390) #394 guards the divisor, not the quotient, so %K = (close−lowest)/(highest−lowest)·100 still overflows when the close sits outside the window, and TA-Lib never checks low ≤ close ≤ high; the repro uses a 1e-303-wide window and a close of 1e6, with no input near 1e306. (2) TA_REAL_MAX (3e37) does not bound this — it constrains optIn real parameters only, and the published domain for a bar is U3's "any finite double", so U3 and "a rejection changes nothing at all" are jointly false as published rather than merely out of domain.
  Why documented rather than fixed. Peek-then-commit roughly doubles composed Update cost on five functions for a corner no feed reaches, with nothing to gate it against. Investigate: batch and streaming tiers return different values (not just a different zero) when the input contains NaN #191 already settled the sibling case — overflow of the library's own arithmetic is a property of double, not of an indicator: not flagged, not gated, not specified. TA_ALLOC_ERR settled the form: name the boundary, then stop. So this became one Part 3 entry, Intermediate overflow, beside Allocation failure; the §2.4 paragraph that enumerated the effects is deleted; the U-tier promise is qualified where it is made; and the public pages carry the fact once (api/README.md §4.2 and the streaming error table).
• TA_StreamAdvance(void *stream) writes through an unchecked void *. It and TA_StreamOutRange are the only two void * parameters in ta_func.h; every other stream call is typed and does diagnose the slip. Passing &s, a double[], or an unrelated struct compiles clean under -Wall -Wextra -Wpedantic -std=c99, returns TA_SUCCESS, and Advance writes 8 bytes into it (measured: caller's handle pointer mutated; an unrelated struct's second field incremented). The write is gated on outRangeCount < TA_MAX_INDEX, so corruption depends on what the second word decodes to. TA_StreamAdvance is brand new in this release — typing it costs nothing now. Tracked in Streaming: type TA_StreamOutRange and TA_StreamAdvance per function (TA_<NAME>_OutRange / TA_<NAME>_Advance) #387 (per-function TA_<NAME>_OutRange / TA_<NAME>_Advance).
• C's OpenAndFill aliasing guard is pointer-equality only. A partially overlapping output passes it, returns TA_SUCCESS, corrupts the fill, and then poisons the ring that memcpys from the already-clobbered input tail — so every later Update is wrong, with no error anywhere. 201 entries, 36 ring-seed sites, one emission site (c_stream.rs:449). C is the only backend with the hole: C# uses Span<T>.Overlaps, and Java/Rust cannot express the call. Both extents are computable at the guard from historyLen and Lookback — the formula C# already uses. Note the guard is currently undocumented, so nothing published is being broken by tightening it. Documented OK (docs/error-handling-spec.md rule N8, Appendix E — decided C: partial buffer overlap is accepted and silently corrupts output #225)
• RSI/CMO under Metastock need Lookback + 2. TA_RSI_Open / TA_CMO_Open reject at exactly Lookback + 1, so the documented pre-flight rule is false for them. Worse, TA_RSI_OpenAndFill on that path returns TA_INSUFFICIENT_HISTORY with outReal[0] already written and outBegIdx/outNBElement set to 13/1 — against the contract that on non-success the buffers are untouched and the indices undefined. Batch tier unaffected. Recorded only in docs/streaming-api-design.md:435-440, which is repo-internal. Removing Metastock (Remove Metastock compatibility (TA_Compatibility) from the TA functions #388) retires the +2 entirely; until then this needs the public exception documented, and the OpenAndFill contract violation fixed regardless — that half is independent of compatibility. No fix needed
• C# SUPERTREND accepts cross-typed aliased outputs (Span<double>/Span<int>, the corpus's only mixed pair) in both tiers and returns silent garbage where C returns TA_BAD_PARAM. C# is not in the v0.8.1 release (IsPackable=false, no NuGet step), so this gates nothing — but it is a real bug and the generator's recorded premise for skipping the check is false. Fix 097b74b
• C# multi-output value structs document the inverse of their real equality. Core_MACD.cs:641-643 says NaN does not equal NaN; measured True — record-struct equality routes through EqualityComparer<double>.Default, so different NaN payloads also compare equal and collapse in a Dictionary. The same docs cite a "Java Value contract" that does not exist. Fix e638e9e

B. Test coverage

• No PR gate runs a single streaming value comparison. pr-codegen-gate.yml has five jobs, all compile-only; its own comment at :226-227 says running ta_regtest "stays in the nightly". The whole stream_verify pass is dev-nightly-only. PR Gate focus on code validity. Numerical validity is heavier and are instead done periodically (nightly CI).
• The Value() vacuity floor is dead in Java and C#. Those servers emit no value_checked/value_legs/value_ok keys, so the floor evaluates 0 > 0 and is skipped. Deleting the Java/C# Value leg tomorrow would be invisible. (C emits 201 of each, Rust 1005/1005/1272.) No fix needed
• The C server has no short-history reject leg — shortHistoryAccepted is 201 in Rust/Java/C#, 0 in ta_codegen_serve.c. server_gen.rs emits it at three sites, none of them C. Done (PR test(serve): C gains the short-history reject leg the other three have (#386) #391, kevinlincg)
• Clone has no generator-side gate in any backend, and only a per-function runtime ratchet in C; in Rust/Java/C# its existence rests on one corpus-wide OR bit. The derivation it needs already exists in peek_suite. Fix 2c1394d (C); No fix needed (Rust/Java/C#)
  Two follow-ups raised against these commits were checked by sabotage and do not hold: 102598205's note that nothing verifies TA_MA_Clone's dispatch switch is true of peek_suite only — stream_verify sends one vector per non-default MAType, so re-routing the EMA arm to TA_SMA_Clone fails the run (exit 79, naming optInMAType:1). And the fork/clone leg is not blind to a shallow-shared buffer despite feeding both handles the same bars: a rolling accumulator reads the trailing slot before overwriting it, so sharing the ring in TA_SMA_Clone fails at the first shared bar (exit 87, clone_bad:"the fork left batch"). The one real gap is that this leg exists in the C server only — 0 clone probes in Rust, Java and C#. No live defect there: every stream handle's copy constructor deep-copies every array field (Java 201 classes / 310 fields / 0 shallow; C# the same; Rust derives Clone over Vec<f64>).
• OpenAndFill's aliasing rejection has no gate in C#, and the four backends' guards have three different strengths. No fix needed
• Nothing verifies the 1,809 new streaming symbols are exported from the Windows DLL — the exact bug class of CHANGELOG Missing TA_GetVersionString function in Windows DLL #57. Fix df228ca

C. Docs

• ta-lib.org/api/stream/ still says "Not yet released — planned for v0.8.x" Fix e63a721. The banner is correct today — CHANGELOG.md still reads ## [0.8.1] Not Released Yet — so the item was never a stale-text bug; it is a release-process gap, and only the C page's banner should go at 0.8.1 (the six Rust/Java/C# ones say "Estimated release: Q1 2027" and are true). Nothing in scripts/ or .github/ removes it, the page is hand-written so regen-check cannot see it, and README-DEVS.md's "After a release" has no step for it — so on release day ta-lib.org would tell every visitor the flagship feature of the release is unreleased. pre-release-checks.py now refuses to cut the release while that one page still carries the banner, beside the "200+ indicators" and soname-reconcile assertions that already have this shape.

D. Release process

• CMake and autotools ship different SONAMEs for the same library, and CMake's changes on every patch release. Fix 7518922 + ABI gate 0e55bb3
• CPack packaging does not follow the SONAME rule. Fix 69191d1 — the .deb installs into /usr but shipped a ta-lib.pc naming /usr/local, so pkg-config pointed every build at a prefix the package never populates (and at a stale /usr/local/lib/libta-lib.so.0 where one exists). package_deb's dist_test_pass was a literal # TODO ... just pretend — the committed digests record "dist_test_pass": "True" for a .deb nothing ever opened; it now requires the payload to carry the soname ABI.manifest declares. And CPACK_PACKAGE_VERSION_PATCH read TA_LIB_VERSION_BUILD, defined nowhere — set(X ${undefined}) unsets X and CPack's own default refilled it, so it was the right value by accident.
  The rest is Post-release: split the .deb into libta-lib1 + ta-lib-dev so the SONAME actually buys co-installability #401 (post-release: the libta-lib1 + ta-lib-dev split renames released artifacts, so it cannot ship in the release it would rename). Multiarch /usr/lib/<triplet> is not being done — ld.so searches the triplet dirs first, so ours is the shadowed copy rather than the shadower, and differing sonames mean nothing collides either way.
• The shipped .so exports every non-static symbol. Spun off as The shipped .so exports 1,519 internal symbols: TA_LIB_API is decorative on ELF #400 — 3,968 exported vs 2,449 header-declared, so 1,519 internals are in the ABI surface, the 401 _Internal seams included. No -fvisibility=hidden anywhere, so TA_LIB_API is decorative on ELF. Hiding them is itself an ABI change, hence its own issue rather than a fix here.
• No abstract/generic streaming tier. TA_FUNC_FLG_STREAM is set on 201 functions and readable via TA_GetFuncInfo, but nothing in ta_abstract.h consumes it — the shipped .so exports 201 each of the nine per-function calls and nothing generic. A wrapper author must hand-write 201 x 9 bindings. Purely additive, closes no door, so it can follow the release. Not planned for v0.8.1
• docs/v0.8.1-checklist branch — 176 of 201 functions, zero boxes ticked in two weeks. Merge, convert, or drop. this is long term work, not blocking on v0.8.1, renamed to not make it release specific.

Refuted — do not re-raise

• "feat(streaming): a rejected Update leaves the handle alone; skipping a bar becomes an explicit call #384 must land before v0.8.1" — done, it is the baseline commit 2ef8a8cfd.
• "Checkpoint-by-re-opening is false for path-dependent functions" — measured false. OBV and EMA-under-Metastock re-open bit-identically; the reproduction retained Lookback+1 bars, which is not what the sentence says to retain.
• "Composed tear is permanent / silently corrupted forever" — it decays to bit-identical, and needs inputs three orders of magnitude past TA_REAL_MAX. Half-refuted, corrected 2026-09-08. "Silently" is right for KC (7,019 bars of wrong values under TA_SUCCESS) and wrong for MACDEXT and STOCHRSI, which reject every later bar forever. "Permanent" is right for those two. And TA_REAL_MAX is not the bound — see the item in section A.

Unverified backlog

64 further candidates were raised and not adversarially checked. Highest-value, roughly ranked: Rust's two naming conventions on one type (Core::SMA beside Core::sma_open) and its opener tuple both recoverable from the handle; arithmetic_overflow blanket-allowed crate-wide in Rust; C reading process-global candle settings live on every Update while the other three snapshot at Open; Java's Core.class at 2.48 MB / 2230 methods; no shared trait/interface over stream handles in any backend; SupertrendOut's fields named real/integer; markdown links rendering literally in 90 javadoc sites; java_stream_suite's emit ratchet stale at 168 against 201; suites building with an empty HelperRegistry for 70 of 201 functions; STREAM FILL VACUOUS's floor being a corpus-wide sum.

[mario4tier]

The 64 candidate findings from this session that were NOT adversarially verified. Recorded here so they survive the session; adversarial verification downgraded 16 of 16 blockers, so treat every line below as a lead to check, not a defect.

Status 2026-09-08: every line in BOTH lists is now struck — fixed, dispositioned, or found stale. The nice-to-have sweep is 19 of 19: 9 fixed, 10 needing no code. Verdicts and reasoning in the comments below.

should-fix (40)

• [C (the hand-rolled PeriodBank Open] TA_MAVP_OpenAndFill leaves *outBegIdx and *outNBElement UNWRITTEN on TA_INSUFFICIENT_HISTORY — the only one of 201 that does Fix e6128e6
• [C header (the behavior itself is c] MAMA's outFAMA is the only declinable out-pointer in the API and the header gives no sign of it — the uniform 201-block header makes it invisible Fix 3a43fde — one sentence in the function's doc block, from the same is_nullable() flag the other three backends already use.
• [C header (all 201 blocks); generat] Every streaming header block points at docs/streaming-api-design.md, which is not installed and not in the release tarball — and Open/Update/Peek/Close carry no header documentation at all Fix cd8720a
• [C (live global read); Rust, Java, ] C streaming reads the PROCESS-GLOBAL candle settings live on every Update and Peek; Rust, Java and C# all snapshot them into the handle at Open — 57 of 61 CDL functions mean different things in C Documented OK — docs/streaming-api-design.md §Concurrency states C's live read as undefined behaviour and calls the managed snapshot an additional protection.
• [Java (outlier); C#, Rust, C; 23 mu] Java is the only backend whose multi-output stream update/peek/value do not RETURN the values — they mutate a caller-owned sink and return void, adding a null-sink rejection class no other backend has No fix needed — the sink is a measured, documented design (Java streaming: multi-output update() allocates a Value record every bar because it is cached for value() #310), and C does not return values either. The one real defect was a stale spec sentence: Fix 4a67789.
• [C# only (C, Java, Rust all expose ] TA_REAL_DEFAULT is internal in C# but public API in C, Java and Rust — every real-parameter XML doc tells the C# user to type the bare literal -4e37 Fix e46e97c — published as Core.REAL_DEFAULT / REAL_MIN / REAL_MAX plus the INTEGER_* trio.
• [C# (Java is already final; Rust's ] Core is the only unsealed public reference type in the C# package; Java's Core is final Fix e46e97c — Core is now sealed.
• [Java, docs, website] Java value(out) on a multi-output stream throws, but both the error spec and the public Java page say value can never fail Fix 4a67789 — the spec's reason clause was the defect; the website row now scopes to value().
• [Java, docs, website] "update allocates nothing" is false in Java for KDJ, STOCHRSI and MA-with-MAMA — the Java page promises "zero bytes per bar" Fix 780cc79
• [Java, C#, docs, website] The peek-allocation story is told three incompatible ways: the design doc names a cause no shipped function has, Java flags three functions, C# flags none, and the Java page says none Fix 780cc79 — the design doc named a cause with zero shipped instances; C# pays neither cost by construction.
• [C#, Java, generator] Every C# multi-output value struct's XML doc asserts a "Java Value contract, which compares bitwise" that does not exist — Java's sink has no equals at all Fix e638e9e
• [Java, Rust, website] The Java and Rust streaming pages end with C's "Discovering streamable functions" paragraph — wrong flag name, wrong link Fix 780cc79 + a84c8a6 — FuncFlags.STREAMING / FuncFlags::STREAM, and the links now point at each page's own anchor.
• [C, Rust, Java, C#, website] MININDEX/MAXINDEX/MINMAXINDEX rebase their index basis at 2^30 in every backend, and only the C# page says so Stale — Remove the 2^30 extrema rebase: rule U4 makes it unreachable (200 emitted sites, 4 backends) #402 removed the rebase and 93b716e had already removed every doc mention. Five stale generator comments went with e46e97c.
• [C, docs, website] TA_INSUFFICIENT_HISTORY, a new public TA_RetCode value, is absent from the CHANGELOG's 0.8.1 section and from the main API page's RetCode table No fix needed
• [C, docs] The error spec says C "mints the handle" in the composed-opener corner; the code returns TA_SUCCESS with *stream left NULL, which breaks the C page's own example Fix 4a67789 — false in BOTH halves, and the arm is measured unreachable.
• [C, Rust, Java, C#, generator] The only all-four-backend Update gate never asserts the increment — it matches the saturation guard's text alone, so "guard, no ++" is green Stale — fixed by 93b716e (U4); deleting the increment at any of the four emitter sites now reddens it by name.
• [C, Java, C#, generator] java_stream_suite, managed_stream_suite and open_core_suite build their subject with an EMPTY HelperRegistry, so for 70 of 201 functions they audit text the generator never emits Fix 59877a3
• [Java, C#, generator] The "value() names the last COMMITTED bar" ordering is pinned in Java only; C# has the identical emitter shape and identical hazard, ungated Fix 756b3d9 — the pin now sweeps Java and C#.
• [C, Rust, Java, C#, generator] Clone has no generator-side gate in any of the four backends, though the derivation it needs already exists in peek_suite Fix 2c1394d (C); No fix needed (Rust/Java/C#)
• [C, Java, generator] open_core_suite's corpus sweeps carry no non-vacuity floor, and java_stream_suite's emit ratchet is stale at 168 against a 201-function corpus Fix fc53b83 (the ratchet); No fix needed for open_core_suite — the narrowing it would catch reddens four sibling assertions in the same cargo test.
• [Java (C# gets it right)] 105 of 201 stream openers document Integer.MIN_VALUE as the default sentinel for functions that have no int parameter — MAMA/SAR/SAREXT and 9 candlesticks take doubles, whose sentinel is -4e37 and is never mentioned Fix e46e97c — now derived from the parameter types. C# had the identical bug in its own stream opener and is fixed with it.
• [Java primarily (Rust and C# have t] No way to drive a stream generically: the 201 handle classes share no supertype, and the metadata tier publishes FuncFlags.STREAMING while ParamHolder can only drive the batch tier No fix needed for v0.8.1 — measured: only 2 of the 5 stream methods are shape-uniform across the corpus (out_range and advance), so a supertype would hold almost nothing a caller could use. The metadata half is the body's No abstract/generic streaming tier, already not planned for v0.8.1.
• [Java (C# has the same field names ] SupertrendOut's public fields are named real and integer — the only sink in the library whose field names carry no meaning Fix e46e97c — renamed outSupertrend / outTrend, free only because SUPERTREND is absent from v0.7.1.
• [Java and C# (Rust is correct — rus] Markdown inline links ship literally into the rendered javadoc — 90 sites across 45 candlestick functions display ([thepatternsite.com](https://...)) as text Fix e46e97c — 152 sites each, not 90; the 6 that remain are inside the welded formula block, tracked in javadoc and XML docs render authored Markdown literally: the Formula block welds prose, and emphasis ships as asterisks #404.
• ~~[Java] Core.class is a 2.48 MB single class with 2230 methods; loading it costs ~55 ms before the first indicator runs, and 82% of its bytecode is the private numerics tiers that nothing outside Core can call~~ Not planned for v0.8.1. No JVM limit is in reach — measured on the built class: constant pool 6,218 of 65,535 (9%), methods 2,229 of 65,535 (3%), largest method 7,252 of 65,535 bytes (11%). So it gates nothing; it is a startup-latency and packaging question for a jar that is not in this release. Worth parking against the Java release rather than dropping: splitting Core is free while Java is unpublished and stops being free the day it ships.
• [C (Windows DLL packaging + CI)] Nothing verifies that the 1,407 new streaming symbols are exported from the Windows DLL — the exact bug class of CHANGELOG Missing TA_GetVersionString function in Windows DLL #57 Fix df228ca
• [Rust/Java/C# packaging, release au] What v0.8.1 actually publishes is stated four incompatible ways, and the release workflows publish only the C assets Fix e63a721
• [C (both build systems, and the shi] CMake and autotools ship different SONAMEs for the same library, and CMake's changes on every patch release Fix 7518922 + ABI gate 0e55bb3
• [release automation / packaging] Housekeeping: remove the old dist/ binaries from git history (3 steps, spread over ~2 releases) #173 Step 2 (get vcpkg off the GitHub tag archive) is scheduled for "the next release" — that is v0.8.1, and nothing in the automation does it Moved to vcpkg release prep for v0.8.1: #173 Step 2, stale overrides, and a pass over the port automation #408 — with the rest of the vcpkg release prep, so it has an owner.
• [C (the only server reporting fill_] STREAM FILL VACUOUS's bar floor is a corpus-wide sum, so 200 of 201 functions can compare zero OpenAndFill bars and still pass Fix 756b3d9 — the floor is per LEG now; the old one allowed 200 of 201 functions to compare zero.
• [Rust, Java, C#] Clone has a per-function runtime ratchet in C only; in Rust, Java and C# the fork leg's existence is proved only by one corpus-wide OR bit No fix needed
• [Rust, Java, C# (57-61 CDL function] Candlestick ring content has no runtime observer at all outside C — the value legs are measured to catch a minority, and the leg that catches all of them is C-only Premise inverted — all four run the full value leg over every candlestick. What is C-only is the complete observer, not an observer.
• [Rust (vs Java/C#); docs] Rust is the only backend whose stream openers return a tuple, and both tuple payloads are exactly recoverable from the handle No fix needed — C hands back the identical two payloads through mandatory out-params, so Rust's tuple is C's contract; Java and C# are the outliers.
• [Rust] One public type carries two naming conventions: Core::SMA / Core::SMA_Lookback beside Core::sma_open / sma_open_and_fill No fix needed — the SCREAMING batch name is the RESULT of 9145c93 (Rust: remaining loose ends before the first crates.io publish #179 C13), and Java and C# carry the same split.
• [Rust (generator: main.rs)] arithmetic_overflow — a deny-by-default correctness lint — is blanket-allowed crate-wide and suppresses exactly zero warnings Fix e46e97c — removed; the crate builds clean. The original rationale was wrong: the lint is const-eval only.
• [Rust (generator: rust_abstract.rs)] FuncFlags::STREAM's rustdoc names a Close method the Rust crate does not have, in C's casing Fix e46e97c
• [docs (describing Rust)] docs/streaming-api-design.md still says the Rust handle "holds its own Core by value"; no Rust stream state holds a Core Fix 780cc79 — measured: zero of 402 handle structs hold a Core, directly or indirectly (Rust optimization: a stream handle clones the whole 280-byte Core, of which 119 of 176 functions read nothing #274/perf(rust): split a stream handle into config/state/OutRange — a step's opt-params are read-only, same shape as #274's Core #276).
• [Rust, Java, C#, C, generator] 20 handles cannot share a type in any of the four backends — no trait, interface, or base class for streams Duplicate of the item above (and the count is 201, not 20) — same disposition.
• [C, generator] The C public header documents none of the streaming contract — 201 blocks that just point at a file make install does not ship Fix cd8720a
• [C, Rust, Java, C#, docs] No documented recovery for a bar committed by mistake; Clone is the mechanism but nothing in the docs connects them or states its cost Fix a84c8a6 — the C and Rust pages regained the checkpoint bullet Java and C# kept. Clone is NOT the mechanism: it forks at the same bar, so it copies the mistake.

nice-to-have (24)

• [C (ta_abstract); generator: stream] The abstract layer advertises TA_FUNC_FLG_STREAM but offers no way to open or drive a stream generically — a working generic dispatcher already exists, private to ta_regtest Duplicate of section D's No abstract/generic streaming tier, already not planned for v0.8.1. The regtest dispatcher covers Open/OpenAndFill/Close only — 3 of 9 verbs, in a noinst header keyed on test-fixture types — so there is nothing to lift.
• [C only (Rust/Java/C# have no way t] "Declinable" stream outputs exist in C only, only for MAMA, and are undocumented in ta_func.h — every other multi-output C function rejects a NULL secondary output Fix 3a43fde
• [C# (same gap in C's ta_abstract, J] The metadata catalogue advertises FunctionFlags.Stream on all 201 functions but offers no way to open one — FunctionCall is batch-only No fix needed — the flag describes the library, not the catalogue's reach, and all four backends ship a batch-only holder tier while setting the bit on all 201. Same decision as the item above.
• [C# (Java has the same split but ca] Two casing conventions for the same indicator inside one class: core.SMA(...) / SMA_Lookback alongside core.SmaOpen(...) / SmaStream / MacdValue No fix needed — the same Rust: remaining loose ends before the first crates.io publish #179 C13 split already dispositioned for Rust. Corrected counts: 603 SCREAMING batch members beside 402 PascalCase stream openers.
• [Java, website] The Java streaming page's error table points at a note that does not exist Fixed — 500942220 deleted the paragraph from all four pages but only Java's pointer survived; "composed" occurred exactly once in website/src/api, in the dangling reference itself.
• [docs, website] The C streaming page still carries a "Not yet released — planned for v0.8.x" banner, and it is the page the 0.8.1 changelog links to Fix e63a721 — the banner is correct today; pre-release-checks.py now blocks the release while it stands.
• [Java, C#, Rust, C, docs] "a copy constructor, which is what every backend already emits" — the two backends the sentence is about emit one that a user cannot call Fixed — measured: Java's copy ctor is package-private, C#'s is internal. Deleted the clause; the load-bearing half (why our clone() sidesteps the Cloneable objection) stays.
• [C, Rust, Java, C#, generator] "update allocates nothing" — the design doc's hard constraint — has no gate in any backend, while peek's weaker version has four Premise wrong — it has gates in all four: a runtime updAlloc accumulator in the Java and C# servers (server_gen.rs:8894-8933), plus Java's corpus-wide exact-map sink pin; is_stateful_call is a generator refusal, stronger than any sweep.
• [Java, C#, generator] no_managed_handle_caches_the_multi_output_value is four hand-picked pairs testing for one literal identifier, but its docstring calls itself a sweep Fixed by deleting it — cachedValue occurs zero times in the generator and all output, git log -S shows C# never had it, and test_java_mama_value_class_protocol already carries the same needle over the same single emitter site.
• [Java (C# deliberately marks the sa] Handle state is package-private, not private, and the jar has no module-info — any class a user drops into package io.github.talib can corrupt a live stream Fixed — every handle field, both constructors and <n>StepImpl are now private, from 10 emitter sites. release=17 means nestmates, so it costs no synthetic accessors. No module-info (inert for classpath consumers) and no sealed package (the jar and a test class dir both contribute the package).
• [Java] The exception family is asymmetric: InsufficientHistoryException has a public constructor while its three siblings' are package-private, none of the four is final, and none accepts a cause Fixed, partly — the three leaf classes are now final (the one-way door: binary-incompatible to add after publication). No cause-accepting constructors: nothing catches-and-rethrows, and they stay additive at any later version.
• [Java] MoneyFlow is dead code that ships in the jar — referenced by nothing in any Java source Fixed — deleted. Stranded by ef2219b7b (ta_abstract: behavioural + coverage parity across C / Java / C# / Rust (umbrella) #164)'s layout move; Java's MFI flattens the struct and never used it.
• [Java (the saturation itself is in ] outRange()'s count silently saturates at MAX_INDEX (100,000,000) — documented on the website, contradicted by the javadoc a Java user actually reads Stale — U4 (93b716e04) fixed both halves. The pre-U4 emitter's own doc comment said "Saturating"; today code, javadoc and website all say update/advance throw past MAX_INDEX.
• [Java] One function's name appears in four different casings on the same class: SMA(...), SMA_Lookback(...), smaOpen(...), Core.SmaStream No fix needed — the same Rust: remaining loose ends before the first crates.io publish #179 C13 decision as the C# item. A unilateral rename reddens registry.rs or fails javac loudly.
• [docs / release process] The docs/v0.8.1-checklist branch is a stale artifact: 176 of 201 functions, zero boxes ticked in two weeks Stale write-up — the de-scope already landed: the file is docs/docs-checklist.md, PR Docs Checklist #247 was renamed to match, and it stands at 34 ticks, not zero. Only the branch name survives the rename.
• [C test infrastructure] Retire ta_ref_serve (pre-cutover oracle) once --xlang-hash covers all languages; generalize ta_064_serve → ta_XXX_serve #116's post-release step — roll the bit-exact release oracle to ta_081_serve — has no owner in the release workflow Fixed — README-DEVS release step (13), conditional on FMA_TRANSITION_TOLERANCE still being 1 so it self-retires. Measured: v0.7.1 had zero fma( in src/ta_func, HEAD has 41, so v0.8.1 is the first release that satisfies the trigger.
• [Rust] The startIdx-anchored _OpenInternal seam is never exercised at runtime in Rust Premise wrong — _open_internal is pub(crate) with no pub re-export, so the only reachable non-zero anchor is a composed cross-call, and those run in every ta_regtest leg.
• [Rust, Java, C# (no leg); C (partia] The peek non-commit twin-handle leg — the only cross-check that a peek stores nothing — is C-only, and it is blind on the extrema family even there Fixed, and it found worse — the generator gate meant to be the complete observer inspected 1 buffer subscript corpus-wide and stayed GREEN with 201 injected peek stores; peek frames bind buffer bases to bare locals (perf(streaming): C peek copies the whole handle struct (~480 B, 174 of 174), the largest per-peek cost left #316) while the detector anchored on sp->. Widened to both spellings (230 sites now) with its own floor; the same hole in Java's and C#'s name_of closed too.
• [Rust] RetCode is the crate's std::error::Error type but carries a Success variant whose Display is "success" No fix needed — Err(RetCode::Success) is unconstructible (all 4,720 Err(...) sites checked), the one path that produced it is fixed in ir_cleanup and gated corpus-wide, and Success stays load-bearing as a value for the _Impl tier and the wire.
• [Rust (Java and C# have the same ga] No shared trait over stream handles: 201 nominal types with identical method shapes and nothing generic to hold them Duplicate of the two above — same disposition.
• [Rust] Index-study outputs are i32 while every index in the Rust API is usize, so the natural round trip does not compile No fix needed — the round trip it says a user cannot write is generated as a runnable doctest on all three index functions and run by cargo test --doc on the PR gate. The usize switch needs a new YAML flag threaded through the store-cast machinery, the stream tier and the abstraction layer.
• [Rust] Multi-output open_and_fill carries an output-aliasing guard that safe Rust makes unreachable, and documents an error a Rust user cannot trigger Fixed — removed. error-handling-spec.md:184 already scored Rust — for S6, so the emitter contradicted the spec, and the guard had already false-rejected (nullable_outputs.rs:105-118, Appendix D item 11). Batch _Impl guard kept.
• [docs (C measured; applies to Rust/] docs/streaming-api-design.md states a Metastock open requirement that the code does not have Moot — Metastock (Remove Metastock compatibility (TA_Compatibility) from the TA functions #388) is fully removed.
• [C, Rust, Java, C#] No cold-start path: an app must build its own per-indicator warm-up buffer before it can Open anything Not planned. Easy to do by caller. Will do if strong demand for it.

[mario4tier]

TA_INSUFFICIENT_HISTORY / CHANGELOG+RetCode-table finding: no fix needed. The streaming page already documents it, and the CHANGELOG's job is to point there, not duplicate it. The "API page's RetCode table" claim was stale — no such table exists outside the stream page.

[mario4tier]

RSI/CMO-under-Metastock-need-Lookback+2 finding: moot. Metastock compatibility (#388) is fully removed — no reference to it remains in RSI/CMO source or docs/streaming-api-design.md — so the +2 exception it produced no longer exists.

[mario4tier]

Landed a real fix from this list (commit e6128e6, pushed to dev): the MAVP-only finding was undercounted -- 14 files across three C-stream emitter sites (12 composed cross-call functions, plus MA's and MAVP's own hand-rolled tiers) returned TA_INSUFFICIENT_HISTORY from OpenAndFill without zeroing outBegIdx/outNBElement. Tightened the existing corpus-wide Open/OpenAndFill sweep (test_open_contract.c, #389) to gate this class specifically, verified red before / green after. Everything else on this list is still open.