fmt: support .Nf / :N / :Nd / :<N placeholder specs (#16c)

[danieljohnmorris]

Summary

Pending entry #16c from `ilo_assessment_feedback.md`. `fmt` used to support only bare `{}` placeholders, which meant agents reaching for "two decimal places" or "right-align to width 5" had to compose `fmt2` + `padl` by hand:

```
fmt "{}%" (fmt2 (* r 100) 1) -- before
fmt "{.1f}%" (* r 100) -- now
```

Token-wise this is the difference between one builtin call per formatted field and three. The spec is intentionally lean: numeric precision, integer width, and string width are the cases that show up in every persona run, and that's all this adds. Zero-pad, sign-prefix, and hex stay out of scope so the surface doesn't grow into Python's `str.format` over time. Agents wanting those still compose via `padl` / `fmt2`.

Repro before / after

Before (`{:.3f}` rejected at verify-time with ILO-T013):
```
fmt "pi={:.3f}" 3.14159 -- ILO-T013: only supports bare {}
```

After:
```
fmt "pi={:.3f}" 3.14159 -- "pi=3.142"
fmt "{:5d}" 42 -- " 42"
fmt "[{:<5}]" "hi" -- "[hi ]"
fmt "{:06d}" 42 -- still ILO-T013 (zero-pad out of scope)
```

What's in the diff

• `src/interpreter/mod.rs`: `Builtin::Fmt` arm now parses each `{...}` placeholder through a new `parse_fmt_spec` + `apply_fmt_spec` pair. Width digits with a leading zero are rejected so `{:06d}` doesn't silently become a 6-wide space pad. Single source of truth for tree, VM, and Cranelift via the existing tree-bridge.
• `src/verify.rs`: literal-template scanner reuses `parse_fmt_spec` for slot counting and unknown-spec rejection. ILO-T013 hint now lists the supported set instead of saying "bare `{}` only".
• `tests/regression_fmt_format_spec.rs`: cross-engine coverage (tree / VM / Cranelift) for each supported spec, the rejection paths (literal `{:06d}`, literal `{:.3}`, computed-template runtime spec), and the lone-brace JSON case.
• `examples/fmt-format-spec.ilo`: in-context demo across the four shapes plus a multi-spec row. Picked up by `tests/examples_engines.rs`, so it's also an engine-level regression.
• `SPEC.md` / `ai.txt`: fmt row updated to match the table that already shipped in `site/.../text.md` (commit 83fd702).

Test plan

• `cargo test --release --features cranelift --test regression_fmt_format_spec` — 36 passed
• `cargo test --release --features cranelift --test examples_engines` — passed
• `cargo test --release --features cranelift` — full suite green except for the pre-existing `skill_md::body_is_thin_bootstrap` failure (also failing on main; SKILL.md is 9548 bytes, cap is 8000, unrelated to this PR)
• `cargo clippy --release --features cranelift --all-targets -- -D warnings` — clean
• `cargo fmt --all -- --check` — clean

Follow-ups

• `skills/ilo/SKILL.md` is already over the 8 KB bootstrap cap on main; needs a separate trim PR.
• If width-padding-char (other than space) becomes a recurring request, the natural extension is `padl s n c` / `padr s n c` rather than adding `{:0<5}` to the spec — keeps the printf surface lean.

[codecov]

❌ 1 Tests Failed:

Tests completed	Failed	Passed	Skipped
4931	1	4930	0

View the full list of 1 ❄️ flaky test(s)

ilo::regression_calendar_arithmetic::add_mo_result_out_of_calendar_range

Flake rate in main: 100.00% (Passed 0 times, Failed 5 times)

Stack Traces | 0.017s run time

thread 'main' (43731) panicked at src/interpreter/mod.rs:2019:21:
attempt to add with overflow
note: run with `RUST_BACKTRACE=1` environment variable to display a backtrace

note: run with `RUST_BACKTRACE=1` environment variable to display a backtrace

To view more test analytics, go to the Test Analytics Dashboard
_{📋 Got 3 mins? Take this short survey to help us improve Test Analytics.}

[danieljohnmorris]

CI status: 3 failing checks (lint, build, coverage), all pre-existing on main:

• lint: skill-token-budget exceeds caps for ilo-builtins-core / -math / -io (unrelated to fmt change)
• build: add_mo_result_out_of_calendar_range panics in calendar-arithmetic code merged in feature: calendar arithmetic builtins #495 (unrelated to fmt change)
• coverage: same calendar-arithmetic panic, plus a version must be X.Y.Z Codecov uploader error
• skill-validate passes

The only check that ran my changes (build) passed every fmt-related test (36 new tests across tree / VM / Cranelift, plus the example file via examples_engines). The single panicking test is in tests/regression_calendar_arithmetic.rs, which I haven't touched.

Confirmed by comparing run 26232609248 (this PR) against run 26232029355 (main HEAD a few minutes ago) — same failure signature, same line numbers, both unrelated to the fmt placeholder work.