feat(pbs): add linked collections and bidirectional snapshots

Implement SDK linked<T> with forward and reverse Cursor snapshots, trusted compiler integration, and neutral FIELDS_PUBLISH producer support. Pin the tested runtime revision and include specs, editor and installed knowledge coverage, conformance fixtures, and DSC-0091 lessons and evidence.
This commit is contained in:
2026-10-23 10:28:44 +01:00
parent c410e35001
commit 77aada6259
38 changed files with 3040 additions and 57 deletions
@@ -0,0 +1,127 @@
# SDK Linked and bidirectional snapshot capture
Owner: `compiler/pbs`. Linked algorithms belong to executable PBS SDK source.
Multi-object publication, storage, faults, ownership, GC and quotas remain generic
compiler/provider mechanisms. This contract builds on [generic structs](generic-structs-and-free-functions.md),
[arrays and optional](typed-arrays-and-optional-execution.md) and the
[shared Cursor protocol](sdk-list-and-cursor.md#cursor-protocol).
## Type, identity and construction
The collections SDK exports `linked<T>`. It is an ordinary invariant generic
struct with stable reference identity: aliases observe mutations of the same
collection. `new linked<T>()` creates an empty collection. Contextual construction
follows generic struct rules, requiring a concrete expected application; constructor
values do not infer T. There is no capacity argument.
A directly optional element, including through grouping, is invalid.
`optional (linked<T>)` remains valid, as do otherwise valid element types containing
optional fields. Values retain their complete groups and normal identities,
including tuples, native values, arrays, functions and contract views. No public
Dynamic, universal Element or wrapper node is introduced.
Storage consists of private doubly linked nodes, each containing one present
value. Logical size equals node count. Empty collections have no head or tail;
one-node collections have the same head and tail. Neighbor links are reciprocal,
forward and backward traversals agree, and there are no extra predecessor or
successor nodes beyond the ends. Nodes and private helpers are not a public API.
## Operations and bounds
| Operation | Result | Behavior |
| --- | --- | --- |
| `size()` | long | Number of elements |
| `first()` / `last()` | optional T | End value, or none when empty |
| `xs[i]` | optional T | Present value at a valid index |
| `xs[i] = value` | unit | Replace a valid position without changing size |
| `add(value)` | unit | Append |
| `addFirst(value)` | unit | Prepend |
| `insert(i, value)` | unit | Insert at `0 <= i <= size` |
| `removeAt(i)` | optional T | Remove and return a valid position |
| `removeFirst()` / `removeLast()` | optional T | Remove and return end value, or none when empty |
| `clear()` | unit | Remove all elements and set size to zero |
| `cursor()` | Cursor<T> | Capture first to last |
| `reverseCursor()` | Cursor<T> | Capture last to first |
Reading, replacing and removing by index require `0 <= i < size`; invalid indices
raise TRAP_OOB even when empty. Insertion accepts size, including zero in an empty
collection. Other insertion indices raise TRAP_OOB. Every valid position contains
an element, so indexed read/removal returns a present optional. Zero and false are
present values. Empty-end results do not suppress type or resource faults.
Insertion/removal changes the logical indices of following elements. Replacement
preserves size. There are no holes, clearAt, capacity, reserve, trim or resize.
Duplicate occurrences are preserved; no equality or uniqueness rule is imposed.
Value search/removal is outside this API, consistently with List.
Size is nonnegative long. Indices accept int and long with localized lossless
promotion, without general numeric conversion or truncation. The SDK checks the
current size against the maximum signed long before incrementing, using the generic
logical guard; an unrepresentable next size is rejected before arithmetic or
mutation. Array's u32 structural maximum is not a Linked size limit. Provider
quotas and applicable structural limits can refuse allocation much earlier.
Receiver/arguments evaluate once in source order, including named application.
## Snapshots and costs
Both capture methods return the existing Cursor<T> contract and preserve its
complete state table. A nonempty cursor starts on its first captured value;
hasNext asks whether a successor exists; next without a successor does nothing.
Only a snapshot born empty has an absent current. Repeated next retains the final
value. Capturing `[A, B, C]` produces `[A, B, C]` or `[C, B, A]` respectively.
Each capture completes before return and owns private immutable singly linked,
non-circular snapshot nodes, independent of the collection's doubly linked nodes.
It uses no snapshot array or live association with editable source nodes.
Subsequent insertion, replacement, removal and clear cannot change the captured
sequence. Shared objects retain identity and later field changes remain visible.
There is no removal/reset through Cursor or public node access.
End operations locate their nodes without traversal. Indexed operations search
from the nearer end and remain O(n) in the worst case. Both captures cost O(n)
work and O(n) additional nodes; next follows one link in O(1). Allocation,
optional-carrier preparation, ownership and eventual GC retain their normal
costs. Repeated indexed reads are not an efficient traversal strategy.
## Complete mutation attempts and lifetime
Before publication, prepare bounds, next size, candidate nodes, optional results,
all field values and call/operand resources. Private generic FIELDS_PUBLISH
publishes complete field groups of source-selected objects after validating all
receivers and groups. Null destinations represent missing neighbors. Algorithms
select the new head/tail/size and neighbor fields in SDK source; the VM has no
Linked opcode or syscall.
No fallible allocation, callback, scheduler or GC boundary may expose partial
publication. A refused mutation preserves all previous head/tail/size values,
elements and neighbor links. This includes optional return preparation for
removals and mutations without new nodes. Individually atomic stores do not prove
whole-operation atomicity. Previously evaluated argument effects are ordinary
source effects and are not rolled back.
Unpublished candidates follow normal ownership and quotas. Capture failure returns
no partial Cursor and does not change the source. Removal and clear release
collection references to discarded nodes and payloads. Doubly linked cycles are
collected by normal tracing when unreachable; no immediate reclamation is promised.
Snapshots retain payloads, not the editable source topology; consumed prefixes
become collectible while the last current value stays retained until released.
## Conformance and distribution
Qualification executes real PBS through common IR/PBX and the production verifier
and VM, with ordinary and optimized producer paths. Cover the full API in empty,
single and multi-node states, reciprocal links and size, nearest-end lookup, all
bounds, int/long and single/named evaluation, complete groups and identity,
optional-element diagnostics, both Cursor orders/states and mutation independence.
Independently observe all original fields, link carriers, nodes and groups around
quota refusals, including node/result/operand preparation. Exercise ordinary GC
safepoints for removed nodes/values, clear, snapshots and prefix/final release.
Generic publication also requires non-PBS producers and malformed/dead operand
coverage. Regress shared Array/List/generic/contract/Cursor behavior and editor/LSP
projections. Ship the specification and executable example with installed SDK
knowledge, compile the actual packaged example outside the checkout and execute
its emitted PBX with the production provider.
See [generic Registry/PBX publication](../../general/specs/typed-registry-and-pbx-v0.md)
for the language-neutral instruction contract.
@@ -152,11 +152,12 @@ faults. List itself rejects directly optional elements.
## Reuse, costs and conformance
The shared implementation is independent of List representation. Future Linked<T>
and Set<T> producers use Cursor<T>; Map<K,V> uses Cursor<keyval<K,V>>. Set/Map gain
The shared implementation is independent of List representation.
[Linked](sdk-linked-and-cursor.md) uses Cursor<T> for forward and reverse captures.
Future Set<T> producers use Cursor<T>; Map<K,V> uses Cursor<keyval<K,V>>. Set/Map gain
no ordering guarantee between captures. A captured keyval association is read-only
and does not follow later association replacement; referenced objects stay shared.
Those collections and keyval are separate deliveries.
Set/Map and keyval are separate deliveries.
For n logical List positions and k present occurrences, capture takes O(n) work
and O(k) additional nodes; next follows one link in O(1). Each node costs allocation,