fix(fasl): replay per-form value echo on a warm load hit (#40)

[pyrex41]

Problem

(load FILE) prints each top-level form's evaluated value (klambda/load.kl → shen.eval-and-print on the raw path, shen.work-through on the tc path) before returning loaded. On a warm FASL-cache hit, boot.lua returned loaded straight out of fasl_replay without re-running the kernel load, so that echo was dropped. (load ...) stdout therefore depended on cache state:

• cross-port divergence — warm shen-lua vs shen-rust / shen-go / shen-cl (which always echo);
• intra-port nondeterminism — shen-lua cold (SHEN_FASL=off) ≠ shen-lua warm.

Fixes #40. Surfaced by Bifrost cases/divergences.json → load-toplevel-echo.

Fix

Capture the echo faithfully rather than re-deriving it:

• During a miss, the per-form echo is the only pr to (stoutput) that fires while inside shen.load-help with rec.in_chunk false. (A form's own (output ...) side effects fire with in_chunk true and already replay by re-running the chunk.) Record those bytes as a new "e" record in stream order, right after each form's "c" chunk.
• On a hit, fasl_replay re-prs each "e" record to (stoutput).

Consequences that fall out for free:

• Interleaving with side-effect output is preserved (echo record follows its chunk in the stream).
• Going through the live pr keeps *hush* gating at replay time — a -q hit stays silent, exactly like a -q cold load. Bytes are captured even under *hush* (the string is computed regardless), so a miss recorded under -q still replays correctly to a non-hushed hit.
• The cosmetic run time: / typechecked in N inferences banners are emitted by load outside shen.load-help, so rec.echoing excludes them — they stay dropped by design (per-run timing noise that would break cross-port comparison).

FASL_FORMAT bumped 3 → 4; stale caches miss and recompile cleanly.

Verification

• Warm == cold == shen-rust == shen-go for the probe, on both the raw and the (tc +) paths.
• -q hit stays silent (matches -q cold).
• SHEN_FASL_DEBUG=1 still shows a hit on the second run — the fast path is intact.
• Kernel tests: 134/134. Port specs: 489/489 (cli_spec gains a load: warm FASL-cache hit drops canonical per-form value echo (cross-port divergence + nondeterminism) #40 regression that drives a private-SHEN_FASL_DIR miss → hit and asserts the value echo returns while the run-time banner stays dropped).

🤖 Generated with Claude Code