chore(experiments): make RESULTS.md a baseline instead of a test output

[pdettori]

experiments/RESULTS.md was both a checked-in artifact and a test output: E2 rewrote it on every run, so any pnpm -r test left the working tree dirty with machine-local wall-clock timings, and the recorded result silently became "whatever the last person to run the suite measured".

This splits the two roles:

• Each run writes its own table to the gitignored experiments/.results/ (SH_E2_RESULTS_DIR=<dir> to redirect).
• The committed RESULTS.md becomes a baseline E2 asserts against, so a change that moves the read counts has to be acknowledged rather than quietly rewriting the record. Refresh deliberately with SH_E2_UPDATE_BASELINE=1 pnpm -C experiments test e2-reconstruction-cost.

Which columns a baseline can actually assert

Measured rather than assumed — E2 run twice locally, diffed against the CI run captured earlier today:

Column	Local	CI	Comparable?
backend / checkpoint entries	53, 203, 1003, 5003 / 6,6,6,6	identical	yes
ratio	8.8, 33.8, 167.2, 833.8	identical	yes
backendBytes	7482, 28508, 140908, 706909	7486, 28512, 140912, 706913	no — consistent +4, serialization is environment-sensitive
checkpointBytes	896, 901, 901, 906	identical	no — same class as above, excluded on principle
backendMs / checkpointMs	vary run to run	—	no

So deterministicView() covers n, both entries columns, and the ratio at the 1 decimal the table carries. Had I compared bytes, this would have gone green locally and failed on CI.

Changes

• experiments/src/report.ts — adds parseE2Table() (reads the baseline back, tolerating surrounding prose and unrelated tables) and deterministicView(), with the measurement rationale documented at the type.
• experiments/test/e2-reconstruction-cost.test.ts — writes to the gitignored dir; compares against the baseline unless SH_E2_UPDATE_BASELINE=1.
• experiments/test/report.test.ts — new, 8 cases: round-trip, prose/other-table tolerance, throwing rather than silently returning nothing on a malformed table, drift-insensitivity, that a real entries change is caught, and that the committed baseline itself parses and round-trips (catching a hand-edit that breaks the shape E2 reads).
• .gitignore, experiments/README.md — the ignored path and the refresh workflow.

Verification

• Three consecutive E2 runs leave RESULTS.md untouched (git diff --quiet clean) while each refreshes the gitignored copy — the actual point of the change.
• Negative control: perturbing a baseline entries cell (6 → 7) fails E2 with a clear diff; restoring it passes and leaves the file byte-identical.
• pnpm -r test: experiments 51 → 59 passed, 0 failures repo-wide. make typecheck clean.
• The baseline refresh in this commit moves only the ms columns — every asserted column is unchanged from what was already committed, which is the evidence the baseline is genuine rather than fitted to my machine.

Note on style

experiments/test/report.test.ts uses double quotes to match the other 12 files in the package. .prettierrc declares singleQuote: true, but nothing in experiments/ follows it and prettier is gated neither in CI nor (working) pre-commit — so I matched the surrounding code rather than leave one odd file out. Worth a separate repo-wide formatting pass if you want the config actually honoured; I didn't fold that into this change.

make lint fails on this machine before and after (Type tag 'typescript' is not recognized — pre-commit's identify is too old). Pre-existing and unrelated.