MemoryPrintk source and native proof

Evidence class: layered source realization, native callback fixtures and pure Nix evaluation. Full matching kernel/dev/system/bundle, selected-header target objects, actual controller preparation, physical output and recovery remain UNVERIFIED. Task 5b.5 remains open. No board/UART/camera or full kernel build was performed.

Worktree /home/jadams/tmp/k230-mainline-uart-progress-memory-printk, branch mainline-uart-progress-memory-printk, base e2a02cb0015e027f10c7d3fbbacfe1534163a329. Owned paths: additive layered patch, recipe/flake variants, native test, this source evidence and only task 5l.1. The coordinator owns review/integration/CI and the sole matching full build.

New exact setup gate k230.uart_progress_memory_printk=1 selects one ordinary pr_info("\nK230_UMK1 s=%u n=%u m=%02x l=%u w=%u\n", ...) in the existing observer. The leading newline gives printk an initial empty message line before the timestamp-prefixed UMK line, separating it from a possible Bash prompt. The branch returns immediately, with no selected final explicit DBCN attempt, retry, force flush, emergency interval or second channel. Existing Memory, progress, CONFIG_RISCV_SBI and actual DBCN availability gates still control observer creation. The new gate alone cannot create an observer. Absent/bare/ invalid values preserve the original observer and worker behavior.

The worker/init tail is byte-identical to the Memory parent. Six 5000ms sleeps, cached snapshots/stop checks, packed release publications, completion, one 45-second wait and one acquire read are unchanged. New flag has ordinary data lifetime; only its setup function is init-only. Observer and fixed format have ordinary lifetimes. Field types are unsigned int; the existing masked fields and bounded six-index loop keep output below 256 bytes. There is no new MMIO, TTY/IRQ access, scheduling policy, lock, firmware or console-driver change. Registered Linux console ownership performs any normal backend operations.

Reproducible native and source checks

PYTHONDONTWRITEBYTECODE=1 TMPDIR=$HOME/tmp \
  python3 -B tests/test_mainline_uart_progress_memory_printk.py
PYTHONDONTWRITEBYTECODE=1 TMPDIR=$HOME/tmp \
  python3 -B -m unittest discover -s tests -p 'test_mainline_uart_progress*.py'
nix-instantiate --parse flake.nix >/dev/null
nix-instantiate --parse nix/kernel-mainline-uart-progress-memory-printk.nix >/dev/null
nix eval --offline --impure --json --expr \
  'import ./docs/evidence/mainline-uart-progress-memory-printk/source-host/check-identities.nix { root = ./.; }' \
  > docs/evidence/mainline-uart-progress-memory-printk/source-host/identities.json
nix build .#kernelMainlineUartProgressMemoryPrintk.src --offline --dry-run --no-link --print-out-paths
flock -n /tmp/k230-nix-build.lock nix build \
  .#kernelMainlineUartProgressMemoryPrintk.src --offline --no-link --print-out-paths

The 10 focused tests compile the actual complete patch chain using native GCC with -Wall -Wextra -Werror, replacing kernel interfaces only with isolated callbacks. They execute selected and original Memory code, asserting exactly one printk/zero selected SBI calls, silent worker and unchanged delays/snapshots, publications/events, all stop boundaries, initialization failure, timeout and terminal-publication race. They test base/Memory/new/config/extension gates, exact setup parsing, invalid-after-valid clearing and ignored printk return without retry/fallback. The 170-test discovery passed before the leading-newline refinement; corrected focused tests passed again afterward. Discovery includes prior source and progress-controller fixtures. These are host fixtures, not concurrent kernel execution or UART reception.

The initial evaluation invocation used an eight-character base revision in a Git flake URL and failed before evaluating identity assertions. The checked-in expression corrects this to the full base revision. No source or artifact failed. The initial no-leading-newline source/evaluation runs were deliberately cancelled when review identified a possible physical-line concatenation after the Bash prompt. The one-call format was corrected to begin with a newline; fresh native/evaluation/source receipts below refer only to that corrected patch. Only the corrected new source derivation was selected by offline dry run and realized under a nonblocking shared lock. No kernel build, fetch or bootstrap was selected. Source path, layered parent, patch, applied worker SHA and identity counts are recorded in proof.json; evaluated paths are in identities.json. The shared lock is released after source realization. All 103 prior package drv identities and 13 exposed kernel source/config pairs matched the full base. New system selects the new kernel, preserves exact parent boot parameters and structured kernel configuration. Existing outputs, including Memory, remain unchanged.

Limits and next gates

Existing selected Memory config and source ground timestamp framing: PRINTK/PRINTK_TIME/8250 console/DW are built in; PRINTK_CALLER is disabled. kernel/printk/printk.c:1355–1394 prints six-decimal time and adds a caller field only under PRINTK_CALLER. PRINTK_EXECUTION_CTX records metadata without forcing a kthread prefix. New actual installed configuration and linked Image must be inspected after the coordinator’s full build; pure configuration equality is not a replacement for that proof. The separately typed controller must retain fresh received arguments/Linux ttyS0 backend markers, zero input, strict new namespace and independent protected recovery; combined nohlt is excluded.

A received UMK record would show progress through observer snapshot/formatting and changed Linux output, not that printk returned or that DBCN caused prior silence. Console ownership/IER handling, nbcon scheduling and possible waits are new output-path dependencies. The kernel wait/call count is not a wall-clock console-return guarantee. Missing output leaves worker/observer/timeouts/ scheduling/output unknown. RX, ordinary root and glass acceptance remain separate.