0.1.4 — stop silently dropping values you wrote in your config

Every fix in this release is one shape of the same bug: a value written in a
config file that the library quietly ignored — the worst failure mode for a
config layer, because it surfaces days later as "why did it train with the
default?". No public API changed.

Highlights

• A nested block is no longer dropped when its type-selector field is left at
  its default. Monitors only ran from __setattr__, so a field kept at its
  default never triggered the swap to the concrete nested Conf; the block
  stayed a bare Conf and every key under it was discarded with only a log
  line. Conf.__init__ now fires monitors, so the nested type is in place
  before the file's keys are applied.
• Parsing is a single, dependency-ordered pass. Scalars are parsed first,
  then containers and nested blocks are parsed into the existing nested Conf,
  repairing only the fields whose type actually changed — instead of settling
  the object and then re-parsing over it, which is where values were being lost.

Correctness fixes

• Rebuilt-field detection holds the object, not its id(). The
  invalidation signature compared nested Confs by address, and CPython
  recycles addresses: a monitor that assigns the same field twice frees the
  original and the second allocation can land on the freed address, so the
  signature came out identical, the repair was skipped, and the file's values
  were lost. Measured 300/300 for that shape (0/300 for a single assignment),
  which is why it stayed hidden — every monitor shipping here is a
  single-assignment chain — but it is a real trap for any same-class
  replacement.
• strict=True now checks fields, not values. It used to re-parse every
  raw file value into the settled object, so a monitor that tightens a
  constraint made strict reject a config that strict=False accepts and gets
  right (n=50 clamped to max 8: fine without strict, ValueError with it).
  Values are already validated as they are parsed; strict is about unknown
  fields, and it now walks the settled objects comparing keys.
• A config list shorter than the field it fills now warns, naming the field
  (a list longer than the field remains an error).

Web / GUI mode (Streamlit)

• Corrections are written back to the right widget. The write-back rebuilt
  each session key from its last path segment only, collapsing every nested
  field onto a top-level one. Through real Streamlit this produced three
  failures: a correction meant for optimizer.lr silently landed on an
  unrelated top-level lr and was emitted in the saved config; a nested derived
  field never converged so the app reran forever; and a list element
  paths.[0] became [0], which the session parser rejects and which then
  re-crashed on every rerun.

Compatibility

• No public API changes; drop-in over 0.1.3.
• Behaviour change worth noting for reproducibility: a config that was
  previously (and silently) falling back to defaults through one of these bugs
  will now apply its written values. If a 0.1.3 run was hitting one of these,
  its effective configuration under 0.1.4 will differ — in the direction of
  what the file actually said.

fix missing dependency

add psutil

add function for dynamic option list

v0.1.2

add function to monitor dynamic option list

GUI is fully working now!

v0.1.1

update readme

The first fully functional version

v0.1.0

update readme