[johnml1135]

Start here: openspec/changes/accurate-work-steps/design.md, then rust/crates/pg-fst/src/work.rs.

Adds the plan for making work_steps the one work unit and retiring steps, and lands the two pieces that are safe today. With a word deadline set, parsing no longer pays a clock read on every work unit: that cost falls from about 90 instructions per unit to 1.6-4.3, or from 3.5-4% of instructions to 0.06-0.19%.

The main question for review is whether a timeout may now fire late. WorkBudget::consume reads the clock once per 64 units, so a timeout can overshoot by at most 63 units. Timeouts are already machine-dependent and non-certifying. The cap check still runs on every reservation, and stopped() always reads the clock, so a poll-only loop still sees the deadline.

• The deadline still acts: a_deadline_stops_reservations_within_one_poll_interval fails with the interval removed (checked by mutation).
• Between polls: a_poll_observes_the_deadline_between_reservation_polls and the_cap_is_exact_between_clock_reads.
• Work counts unchanged: identical work_steps with and without a deadline on Indonesian and Amharic (1,613,681 and 32,238,244).
• Rename: SynthesisBudget::steps() returned the work total, so it is now work_steps(). No behaviour change.

Not done here: steps still exists everywhere, and no charge, cap, default, batch TSV column or ranking changes. Those are later task groups behind stated gates, including a recall gate before the attempt cap goes. The A/B test of whether a change in work_steps predicts a change in instructions has not run, so D5's thresholds are provisional.

Base 71245076, head 14eac325. pg.ps1 -Mode test: 2,244 passed, 66 skipped (before the last spec-only amend). -Mode preland: passed, with conformance skipped because no gated path changed. A forced -All run failed only that driver, because hc-conformance.exe for the pin is not built here. No parse comparison against a base build was run on real grammars.

Next: land the measurement tools (task group 2), then run the A/B set.

---

Evidence behind the plan

Measured on native Windows (i7-12700) over five real grammars, with PMC on context switch. Instructions repeat to 0.15-0.40% per word, core cycles to 5-17%. Instructions per step span 52x and per work step 2.2x. One per-grammar constant predicts a word's instructions to a median 3-14%.

Valgrind in WSL agrees with PMC at per-word log R² 0.995-0.999. Size-aware counters cut held-out median error from 18.6% to 9.0%, using bytes allocated and FST registers copied. Details and caveats are in design.md. Raw data contains real-language words and stays local.

Paths not taken

• Cycles as the unit: not deterministic, and no unprivileged Windows counter.
• A fitted per-class cost table: weights vary up to 18x across grammars and do no better than one per-grammar constant.
• Moving Score::key to work now: a plain swap shrinks raw_paths about 30x and reopens a recorded mis-ranking, so it waits for a measured weight (D7).

🤖 Generated with Claude Code