Shared request DValue

Changes

src/wasm/core.cpp creates one request DValue on the first handler call. It preserves the root pointer until request reset.
The core refreshes props, call, unit, out.status, out.headers, out.cookies, session, websocket, and connection before each handler.
The core preserves all application keys. It restores the caller's props and unit values after nested handler calls.
The core reads the shared out value after each handler. It applies status, headers, cookies, and the compatible buffer field.
The core removed the handler input release hook and per-frame input tracking. Request reset frees the shared handle and payload.
src/capy/compiler.cpp imports bearer_handler_input_dval. The handler prologue passes its borrowed pointer without a retain or release.
src/wasm/abi.h sets the core ABI to 34 and the compiler unit ABI to 36.
site/tests/capy-request-writeback.capy proves shared application keys and nested response writes. The entry writes request.user. The component reads it and writes request.pagefacts. The entry reads that key after the call.
scripts/test_capy_phase1.sh checks the shared values, nested response changes, restored props, and the new borrowed-parameter ARC counts.
scripts/test_capy_native_compiler.cpp and scripts/test_capy_native.sh check the new import.
scripts/capy_artifact_golden.sha256 records the inspected ABI-wide artifact changes.
docs/capy-language.md specifies the shared request exception and the Bearer-owned keys.
The request type page, guides 9 and 10, handler pages, unit-call page, and context API pages now describe request keys.
The unit_context() and unit_context_set() pages mark both APIs as deprecated. They name application request keys as the replacement.
skills/capy-bearer/SKILL.md now teaches shared request keys. The installed skill copy also matches it.
scripts/generate_capy_docs.py regenerated the affected files under site/doc/content/ and the documentation search index.

Concurrent work also changed docs/deploy.md and files under editors/vscode/. This task did not edit those source files.

Measurement

The prior recorded result was 1986 microseconds per component call.

The new result was 934.3 microseconds per call. The benchmark made 1,000 component calls in one CLI request. The request contained a 200-key request.application map. The component read key 199 on each call. Fifteen samples ranged from 903.6 to 1041.4 microseconds per call. The median improved by 53 percent.

Gates

Passed:

bash scripts/build_core_wasm.sh
bash scripts/build_linux.sh
bash scripts/build_capy.sh
python3 scripts/generate_capy_doc_signatures.py --capyc bin/capyc
python3 scripts/generate_capy_docs.py
bash scripts/test_capy_native.sh
bash scripts/test_capy_artifact_golden.sh after diff inspection
bash scripts/test_capy_artifact_golden.sh --write
bash scripts/test_capy_artifact_golden.sh
bash scripts/test_capy_reproducible_artifact.sh
bin/capyc --check site/
bin/capyc --format --check site/
Bearer restart and scripts/systemd/wait-ready.sh
bash scripts/test_capy_phase1.sh
bash scripts/test_capy_guide_examples.sh
python3 scripts/check_capy_doc_examples.py
python3 scripts/test_capy_doc_generators.py
bash scripts/test_capy_reference.sh
The isolated native CLI runner group
bash scripts/test_wasm_core_smoke.sh

Phase 1 failed while the fixture still expected the old header value. The final run passed after the fixture checked the component's header value.

Not completed

scripts/run_cli_tests.sh does not expose the requested http, site, security, component, and task groups separately. Its --list output contains only native. I ran that isolated group. I did not run the complete service-changing suite.
The ludum.live clone failed. Deploy key 31 on this host cannot read udo/ludum.live. No ludum source was available for capyc --check --source-root.

No commit or push was made.

Live value representation

The shared request handle now points directly to its core-heap DValue. Refresh and writeback update that object in place. They do not pass request state through BRRB. Request reset deletes the value and frees its 28-byte handle.

The 1,000-call, 200-key component benchmark now has a median of 39,524 microseconds per request, or 39.5 microseconds per call. Five samples ranged from 34,070 to 42,779 microseconds per request. The previous live-request result was 934.3 microseconds per call.

4.5 KiB Raw Blame History